--- Log opened Fri May 25 00:00:41 2012 | ||
-!- puffin444 [62e3926e@gateway/web/freenode/ip.98.227.146.110] has joined #shogun | 00:08 | |
shogun-buildbot | build #551 of java_modular is complete: Failure [failed test_1] Build details are at http://www.shogun-toolbox.org/buildbot/builders/java_modular/builds/551 blamelist: sonne@debian.org | 00:20 |
---|---|---|
-!- puffin444 [62e3926e@gateway/web/freenode/ip.98.227.146.110] has quit [Quit: Page closed] | 01:38 | |
-!- shogun-buildbot [~shogun-bu@7nn.de] has quit [Ping timeout: 252 seconds] | 01:49 | |
-!- shogun-buildbot [~shogun-bu@7nn.de] has joined #shogun | 01:49 | |
CIA-113 | shogun: Sergey Lisitsyn master * rede11ce / src/shogun/classifier/svm/SVMOcas.cpp : Made SVMOcas support subsets - http://git.io/27jFxw | 01:52 |
CIA-113 | shogun: Sergey Lisitsyn master * r37f1841 / (2 files in 2 dirs): Removed a few warnings - http://git.io/Q2r26g | 01:52 |
-!- av3ngr [av3ngr@nat/redhat/x-zaxxfumfvxtiubyv] has joined #shogun | 02:10 | |
-!- blackburn [~blackburn@188.122.250.167] has quit [Quit: Leaving.] | 02:11 | |
-!- n4nd0 [~nando@s83-179-44-135.cust.tele2.se] has quit [Quit: leaving] | 02:17 | |
-!- Netsplit *.net <-> *.split quits: gsomix | 02:38 | |
-!- Netsplit *.net <-> *.split quits: @sonney2k, sonne|work, av3ngr | 02:39 | |
-!- Netsplit over, joins: gsomix | 02:42 | |
-!- av3ngr [av3ngr@nat/redhat/x-zaxxfumfvxtiubyv] has joined #shogun | 02:44 | |
-!- sonne|work [~sonnenbu@194.78.35.195] has joined #shogun | 02:44 | |
-!- sonney2k [~shogun@7nn.de] has joined #shogun | 02:44 | |
-!- ServerMode/#shogun [+o sonney2k] by pratchett.freenode.net | 02:44 | |
-!- Netsplit *.net <-> *.split quits: gsomix | 02:45 | |
-!- Netsplit over, joins: gsomix | 02:51 | |
-!- vikram360 [~vikram360@117.192.170.193] has joined #shogun | 03:01 | |
-!- vikram360 [~vikram360@117.192.170.193] has quit [Read error: Connection reset by peer] | 03:50 | |
-!- gsomix [~gsomix@83.234.54.21] has quit [Read error: Operation timed out] | 06:24 | |
-!- gsomix [~gsomix@83.234.54.21] has joined #shogun | 06:24 | |
-!- gsomix [~gsomix@83.234.54.21] has quit [Quit: Ex-Chat] | 07:24 | |
-!- pluskid [~pluskid@li400-235.members.linode.com] has joined #shogun | 07:43 | |
-!- mode/#shogun [-bb *!*46e7e626@*.70.231.230.38 sonney2k!~shogun@7nn.de] by sonney2k | 07:55 | |
-!- uricamic [~uricamic@2001:718:2:1634:5c56:9bc4:b0a1:c02b] has joined #shogun | 08:49 | |
-!- pluskid [~pluskid@li400-235.members.linode.com] has quit [Ping timeout: 244 seconds] | 09:38 | |
-!- av3ngr [av3ngr@nat/redhat/x-zaxxfumfvxtiubyv] has quit [Quit: That's all folks!] | 09:55 | |
-!- pluskid [~pluskid@111.120.53.76] has joined #shogun | 09:56 | |
-!- pluskid [~pluskid@111.120.53.76] has quit [Ping timeout: 252 seconds] | 10:33 | |
-!- pluskid [~pluskid@li164-218.members.linode.com] has joined #shogun | 10:33 | |
-!- pluskid [~pluskid@li164-218.members.linode.com] has quit [Ping timeout: 240 seconds] | 10:42 | |
-!- eric___ [2e1fd566@gateway/web/freenode/ip.46.31.213.102] has joined #shogun | 10:44 | |
eric___ | hi all | 10:44 |
eric___ | sonne|work: is the new multiclasslabels interfaces working ? | 10:44 |
sonne|work | eric___: should be | 10:53 |
eric___ | eric___: I will test it today, thx | 10:55 |
eric___ | sonne|work: btw, I have mentioned few days ago, smthg strange with the multiclass Xvalid: for some kernel machine, the Xval mean is always equal to 0.1. When I split my data manually several times, I got mean > 0.7* | 10:57 |
sonne|work | eric___: can you give us some example to reproduce the problem? | 10:58 |
-!- pluskid [~pluskid@111.120.53.76] has joined #shogun | 10:59 | |
eric___ | sure. how can I send you the features ? | 10:59 |
sonne|work | eric___: how big is this? can you reproduce this on some small example (the smaller the better, e.g. random/artificial data) | 11:03 |
eric___ | sonne|work: okay, I will try to reproduce the problem. I am currently working with visual data so it's pretty big. | 11:11 |
sonne|work | eric___: thanks - that will make it much easier to fix the problem | 11:15 |
-!- n4nd0 [~nando@s83-179-44-135.cust.tele2.se] has joined #shogun | 11:17 | |
sonne|work | n4nd0, pluskid have you seen - we are drowning in warnings http://shogun-toolbox.org/buildbot/builders/java_modular/builds/553/steps/compile/logs/warnings%20%28213%29 | 11:22 |
sonne|work | if you have time please fix the ones in your code | 11:22 |
pluskid | sure | 11:24 |
eric___ | No more "get_label" method in CLabels ? | 11:35 |
-!- alexlistens [6d2d0022@gateway/web/freenode/ip.109.45.0.34] has joined #shogun | 11:36 | |
sonne|work | eric___: exactly | 11:36 |
sonne|work | eric___: each machine now returns a particular label object, like CMulticlassLabels (for multiclass machines) | 11:37 |
sonne|work | etc | 11:37 |
sonne|work | so you will need to do CMulticlassLabels* mc_labels=CMulticlassMachine::obtain_from_generic(labels); | 11:38 |
sonne|work | mc_labels.get_label() | 11:38 |
-!- nicococo [~nico@lacedcoffee.ml.tu-berlin.de] has joined #shogun | 11:38 | |
eric___ | sonne|work: perfect | 11:39 |
-!- n4nd0 [~nando@s83-179-44-135.cust.tele2.se] has quit [Quit: leaving] | 11:43 | |
-!- n4nd0 [~nando@s83-179-44-135.cust.tele2.se] has joined #shogun | 11:43 | |
-!- ckwidmer [~chris@HSI-KBW-046-005-237-106.hsi8.kabel-badenwuerttemberg.de] has joined #shogun | 11:45 | |
n4nd0 | nicococo: hey! I started yesterday to write some code to input the problem to mosek | 11:47 |
-!- karlnapf [~heiko@host86-171-220-235.range86-171.btcentralplus.com] has joined #shogun | 11:48 | |
karlnapf | sonney2k, in 20 mins it is 10 UTC not 11 UTC | 11:48 |
karlnapf | I thought we would meet at 11 UTC? | 11:48 |
karlnapf | sonne|work ^ | 11:48 |
-!- vojtech [9320543b@gateway/web/freenode/ip.147.32.84.59] has joined #shogun | 11:58 | |
vojtech | sonne: hi Soeren, is the meeting at 10:00UTC or 11:00UTC ? | 11:59 |
-!- blackburn [~blackburn@188.122.250.167] has joined #shogun | 12:04 | |
eric___ | sonne|work: I created CMulticlassLabels and trained mc_svm. But I cannot write CMulticlassLabels* training_output = mc_svm->apply(training_features); because "error: invalid conversion from 'shogun::CLabels*' to 'shogun::CMulticlassLabels*' [-fpermissive]" I have to write CLabels* training_output = mc_svm->apply(training_features); Is that what you wwant ? | 12:05 |
eric___ | sonne|work: okay I find out .. Have to use "apply_multiclass" has we discuss last time | 12:06 |
sonne|work | 11 UTC | 12:06 |
sonne|work | sry | 12:06 |
blackburn | eric___: you may use casting there | 12:06 |
eric___ | blackburn: yes that's what I feared, casting casting .. :) | 12:07 |
eric___ | gj btw | 12:07 |
sonne|work | eric___: or you use the helper directly CMulticlassLabels* mc_labels=CMulticlassMachine::obtain_from_generic(mc_svm->apply()); | 12:09 |
blackburn | pluskid: I have crash here with ECOC encoding decoding | 12:11 |
blackburn | let me try to come up with reproducing example | 12:11 |
blackburn | pluskid: ok easy | 12:12 |
blackburn | just train and apply machine two times | 12:12 |
eric___ | I have crashed LibSVM model->nr_class=9 while num_classes=10, never happened before with simple CLabels | 12:17 |
eric___ | going for lunch, see you. | 12:17 |
sonne|work | eric___: assertion or what? | 12:18 |
sonne|work | I found many mistakes in examples simply because we didn't have checks :) | 12:18 |
pluskid | blackburn: with any ECOC strategy? | 12:21 |
-!- karlnapf1 [~heiko@host86-185-113-34.range86-185.btcentralplus.com] has joined #shogun | 12:21 | |
-!- blackburn [~blackburn@188.122.250.167] has quit [Read error: Connection reset by peer] | 12:22 | |
-!- alexlistens [6d2d0022@gateway/web/freenode/ip.109.45.0.34] has quit [Ping timeout: 245 seconds] | 12:22 | |
-!- karlnapf [~heiko@host86-171-220-235.range86-171.btcentralplus.com] has quit [Ping timeout: 245 seconds] | 12:23 | |
-!- blackburn [~blackburn@31.28.59.65] has joined #shogun | 12:23 | |
karlnapf1 | tets | 12:25 |
sonne|work | karlnapf1: tuts | 12:27 |
karlnapf1 | sonne|work :D | 12:27 |
karlnapf1 | just checking my connection | 12:27 |
sonne|work | 20% package loss | 12:27 |
karlnapf1 | annoying | 12:28 |
karlnapf1 | I think Ill go somewhere else next week during the day | 12:28 |
karlnapf1 | all the workshops and coffee places here got better connections than mine at home | 12:29 |
sonne|work | lets hope my mobile phone will work :) | 12:29 |
karlnapf1 | sonne|work are you IRCing via phone? | 12:31 |
karlnapf1 | nice | 12:31 |
sonne|work | karlnapf1: notebook -> mobile phone | 12:31 |
karlnapf1 | ah alright | 12:31 |
sonne|work | btw pluskid that is how I am only in trains | 12:32 |
sonne|work | now way that german trains have wlan :) | 12:32 |
pluskid | sonne|work: oh, I see! :D | 12:32 |
sonne|work | well some ICEs have but... | 12:32 |
karlnapf1 | sonne|work the public transport in Berlin hasnt, its too poor ;) | 12:32 |
sonne|work | yeah | 12:32 |
sonne|work | I am already happy if trains go at all | 12:33 |
karlnapf1 | London underground has neither, and there is also no mobile internet available there | 12:33 |
karlnapf1 | hehe | 12:33 |
sonne|work | karlnapf1: I've just seen this nice graph http://www.ohloh.net/p/shogun | 12:33 |
sonne|work | with the mouse cursor hover over F | 12:33 |
karlnapf1 | I keep adapting to the trains here: If you have to wait more than 5 mins you get angry | 12:34 |
karlnapf1 | sonne|work yes, I already knew that one. its nice :) | 12:34 |
karlnapf1 | . | 12:35 |
pluskid | sonne|work: git says I have changes to commit in the data directory | 12:36 |
pluskid | and git diff shows this: http://pastebin.com/BqfhYZRw | 12:36 |
pluskid | do you know what's this? | 12:37 |
sonne|work | pluskid: you have to do git submodule update | 12:37 |
sonne|work | pluskid: we have the data in a seperate repository | 12:37 |
sonne|work | separate | 12:37 |
pluskid | sonne|work: ah, works, thank you | 12:37 |
blackburn | pluskid: will you take a look? | 12:38 |
sonne|work | blackburn: you didn't really give more details... | 12:38 |
pluskid | blackburn: yeah, after I fix the doc warnings | 12:38 |
blackburn | just train multiclass machine two times | 12:38 |
blackburn | BOOM! | 12:39 |
blackburn | train and apply | 12:39 |
blackburn | for some reason submachine is null | 12:41 |
CIA-113 | shogun: Chiyuan Zhang master * r93b9050 / (14 files in 4 dirs): Fix missing doc warnings and correct some label types in MC Strategies. - http://git.io/n1-yQw | 12:43 |
CIA-113 | shogun: Soeren Sonnenburg master * r6c4cdd5 / (14 files in 4 dirs): Merge pull request #552 from pluskid/multiclass - http://git.io/cSPxHw | 12:43 |
blackburn | hahah | 12:43 |
blackburn | sonne|work: one more reason to use *STANDART* data structures | 12:44 |
pluskid | blackburn: you find the bug? | 12:44 |
blackburn | yes | 12:45 |
pluskid | what's it? | 12:45 |
blackburn | clear_array | 12:45 |
blackburn | instead of reset_array | 12:45 |
pluskid | hmm... where? | 12:45 |
blackburn | in train of multiclass machine | 12:45 |
pluskid | btw: you are super fast :D | 12:45 |
pluskid | I haven't finished re-compiling yet :D | 12:46 |
blackburn | ccache :D | 12:46 |
pluskid | already using ccache+clang | 12:46 |
blackburn | hmm | 12:46 |
pluskid | don't know whether using ccache-swig might helps further | 12:46 |
pluskid | already much faster than before | 12:46 |
blackburn | full recompile takes 3-4 mins here | 12:46 |
pluskid | with ccache? | 12:47 |
blackburn | yes | 12:47 |
pluskid | hmm, then I'm not too slower than yours :D | 12:47 |
blackburn | not slower at all I think | 12:48 |
blackburn | cause I am using gcc | 12:48 |
blackburn | afaik it is slower? | 12:48 |
pluskid | blackburn: btw, will you fix it? | 12:48 |
pluskid | gcc uses more memory | 12:48 |
CIA-113 | shogun: Sergey Lisitsyn master * r19ddc93 / (2 files): Fixed wrong handling of submachines array in multiclass machine - http://git.io/a1TchQ | 12:48 |
blackburn | fixed what?:D | 12:48 |
pluskid | haha | 12:48 |
pluskid | super fast | 12:49 |
-!- oliver [c07c1afb@gateway/web/freenode/ip.192.124.26.251] has joined #shogun | 12:51 | |
-!- karlnapf1 [~heiko@host86-185-113-34.range86-185.btcentralplus.com] has quit [Quit: Leaving.] | 12:54 | |
-!- heiko [~heiko@host86-185-113-34.range86-185.btcentralplus.com] has joined #shogun | 12:55 | |
@sonney2k | blackburn, the PolarBearAttack msg me with "learn not to ban hackers" | 12:55 |
@sonney2k | cool or? | 12:55 |
blackburn | :D | 12:55 |
blackburn | are you trying to convince me it is a bot? | 12:56 |
@sonney2k | that bot was even better than betty | 12:56 |
blackburn | I believe but you would have to agree it is a very cool and smart bot | 12:56 |
blackburn | :D | 12:56 |
@sonney2k | yeah it is | 12:56 |
@sonney2k | pluskid if you know of a way to split up a >10 MB .cpp file and to compile the splits separately but link them together in the end | 12:58 |
@sonney2k | that would speed it up! | 12:58 |
blackburn | sonney2k: should ocas be faster than liblinear (binary)? | 12:58 |
@sonney2k | depends on problem | 12:58 |
@sonney2k | low dim problems - maybe | 12:58 |
pluskid | sonney2k: can't we compile different modules separately? (Evaluation, Classifier, etc.) | 12:58 |
@sonney2k | but ask vojtech :D | 12:58 |
blackburn | oh meeting in a min | 12:59 |
blackburn | :D | 12:59 |
@sonney2k | pluskid, yeah we had that - but this was causing lots of grief because modules needed to know types from other modules | 12:59 |
pluskid | hmm... | 12:59 |
@sonney2k | pluskid, example is: you need Labels in evaluation | 13:00 |
@sonney2k | kernels need features | 13:00 |
@sonney2k | machines need features | 13:00 |
@sonney2k | etc | 13:00 |
@sonney2k | so now we only have shogun/modshogun left | 13:00 |
-!- puffin444 [62e3926e@gateway/web/freenode/ip.98.227.146.110] has joined #shogun | 13:01 | |
pluskid | I'm not sure, but as long as header files are included properly, they can be compiled separately, aren't they? | 13:01 |
@sonney2k | pluskid, but a general tool must exist to split up .cpp files | 13:01 |
pluskid | each module into an .o file, and then link together | 13:01 |
@sonney2k | I mean it is not that difficult | 13:01 |
pluskid | hmm... | 13:01 |
@sonney2k | pluskid, it is not libshogun that is big | 13:01 |
@sonney2k | libshogun compiles in a few sec | 13:02 |
@sonney2k | it is the interfaces | 13:02 |
@sonney2k | hmmhh so who is missing | 13:02 |
-!- cheng [73406f11@gateway/web/freenode/ip.115.64.111.17] has joined #shogun | 13:02 | |
@sonney2k | cheng | 13:02 |
@sonney2k | welcome | 13:02 |
pluskid | yeah, I see swig takes a long time | 13:02 |
@sonney2k | (only swig) | 13:02 |
@sonney2k | heiko - is arthur joining? | 13:02 |
@sonney2k | does anyone know where wiking is / alex binder? | 13:02 |
cheng | Hi sorry to be a bit late | 13:03 |
heiko | sonney2k, I am not sure, I wrote him another email, but he is always very busy during the day | 13:03 |
@sonney2k | nicococo, ? | 13:03 |
pluskid | will discuss about this compiling stuff later | 13:03 |
@sonney2k | nips deadline yeah | 13:03 |
heiko | sonney2k, yes, I am involved this time aaaah :) | 13:03 |
@sonney2k | very good | 13:03 |
@sonney2k | cite shogun in you paper | 13:03 |
@sonney2k | blackburn, you forgot that last time | 13:03 |
blackburn | sonney2k: what? | 13:04 |
blackburn | what did I forgot? | 13:04 |
pluskid | cite shogun | 13:04 |
@sonney2k | so lets wait 1 more minute | 13:04 |
@sonney2k | not good that wiking / alex are missing | 13:04 |
@sonney2k | let me write them an email | 13:04 |
-!- alexlistensagain [c25fae8d@gateway/web/freenode/ip.194.95.174.141] has joined #shogun | 13:05 | |
@sonney2k | alexlistensagain, ah hey | 13:05 |
alexlistensagain | the captchas in the webchat are very hard gfor me to identify ^^ | 13:05 |
@sonney2k | alexlistensagain, any idea where wiking is? | 13:05 |
n4nd0 | :D damned captchas | 13:05 |
blackburn | sonney2k: I did not forget to cite shogun | 13:05 |
@sonney2k | ok then | 13:05 |
alexlistensagain | I will ask viking via mail | 13:06 |
@sonney2k | I already asked | 13:06 |
alexlistensagain | maybe he is eating now: 1 PM | 13:06 |
@sonney2k | I know gsomix is not joining | 13:06 |
@sonney2k | well date is known for 1 month... | 13:06 |
@sonney2k | anyways | 13:06 |
@sonney2k | lets start | 13:06 |
@sonney2k | hello and welcome everyone to our second meeting. | 13:06 |
@sonney2k | Before we continue - nico goernitz (nicococo) and alexander binder (alexlistensagain) are here today too. So please introduce yourself. Nico care to start? | 13:07 |
@sonney2k | alexlistensagain, could you then? seems nico is asleep. | 13:07 |
alexlistensagain | I am Alex | 13:07 |
alexlistensagain | I am advising Viking (Viktor) | 13:08 |
alexlistensagain | on a project about latent structures SVM | 13:08 |
alexlistensagain | I am doing machine learning and computer vision at TU Berlin | 13:08 |
alexlistensagain | near the end of my PhD (some 120 oages written ;)) | 13:08 |
@sonney2k | congrats! | 13:08 |
alexlistensagain | if you want something from me pls kick me | 13:08 |
@sonney2k | thank alexlistensagain | 13:09 |
alexlistensagain | best is a string of nagging mails | 13:09 |
@sonney2k | nicococo, alive now? | 13:09 |
blackburn | alexlistensagain: are you familiar with HOG btw? | 13:09 |
alexlistensagain | yes | 13:09 |
@sonney2k | if not then let me give you an update over what happened to shogun in the last 4 weeks | 13:09 |
blackburn | alexlistensagain: hmmm can I ask you some Qs after the meeting? | 13:09 |
@sonney2k | With the help of pluskid, n4nd0, wiking, blackburn, heiko and gsomix we made automagic memory (de)-allocations in SGVector, Matrix etc possible. | 13:09 |
alexlistensagain | yes, you can do | 13:10 |
@sonney2k | One can now just write x=SGVector<float64_t>(10) to get a vector of length 10 and does not have to take care of de-allocating memory. | 13:10 |
@sonney2k | y=x and all that are zero-copy operations and ref-counting is being taken care of underneath. | 13:10 |
@sonney2k | So thanks to all for the help in getting this pretty nice system done :-) We still have to convert SGString/List and SGSparseMatrix/Vector to this system but this will be much less intrusive. | 13:10 |
@sonney2k | However... | 13:10 |
@sonney2k | The probably most intrusive change is that we went from normal double* which machines returned as output to more complex Label objects | 13:11 |
@sonney2k | BinaryLabels, MulticlassLabels, RegressionLabels, StructuredOutputLabels,... | 13:11 |
@sonney2k | E.g binary labels are now just sign(f(x) aka +1/-1 and have f(x) stored as 'confidences') | 13:11 |
@sonney2k | I also want to thank students for staying around in #shogun - only due to that we could quickly render decisions in joint discussions about how things should develop. | 13:11 |
@sonney2k | some could be a bit more active / communicative though but hey | 13:11 |
@sonney2k | so any questions to that? | 13:12 |
@sonney2k | 1 | 13:12 |
@sonney2k | 2 | 13:12 |
@sonney2k | 3 | 13:12 |
cheng | like! | 13:12 |
@sonney2k | 4 | 13:12 |
@sonney2k | ...1 | 13:12 |
@sonney2k | 0 | 13:12 |
@sonney2k | ok | 13:12 |
blackburn | everybody is asleep | 13:12 |
alexlistensagain | I liked it as well | 13:12 |
n4nd0 | awake here :P | 13:12 |
@sonney2k | then lets continue | 13:12 |
alexlistensagain | we ould have needed labels w structrue anyway | 13:12 |
@sonney2k | OK, so GSoC is now officially running since the beginning of this week and you know that you are on track if you (your student) has already committed the first patches related to his work. Recall mid-term evaluations are already due on July 13 - so by then the first part of the work should be done. | 13:12 |
vojtech | sonney2k: does it mean you defined types for output spaces? | 13:13 |
@sonney2k | vojtech, better discuss that with nicococo and n4nd0 | 13:13 |
@sonney2k | afterwards | 13:13 |
vojtech | ok | 13:13 |
@sonney2k | vojtech, but more or less ... coffin for SO | 13:13 |
vojtech | I see | 13:14 |
@sonney2k | back to topic | 13:14 |
n4nd0 | vojtech: that's the plan but it is not looking yet like that :S | 13:14 |
@sonney2k | Please note that we don't want one big patch at the end of mid-term/GSoC but many small ones continuously floating in. | 13:14 |
n4nd0 | vojtech: we can talk about it later if so | 13:14 |
blackburn | it looks like everybody is waiting for the afterparty | 13:14 |
@sonney2k | So maybe each student/mentor combination can quickly state where they are | 13:14 |
@sonney2k | what they have been doing and problems they see (of any nature). | 13:14 |
@sonney2k | Again if someone is not happy with the work say so either here or to us in private and we see what we can do. | 13:14 |
alexlistensagain | viking started to code something | 13:15 |
@sonney2k | alexlistensagain, something sounds like something :) | 13:15 |
@sonney2k | I guess with wiking missing you should summarize later | 13:15 |
alexlistensagain | I did not scrutinize it but I have no complaints now | 13:15 |
@sonney2k | since everyone has to write short 'I did that' summaries starting from monday to the mailinglist ... I hope we will find out more details | 13:16 |
@sonney2k | vojtech, uricamic : want to continue? | 13:16 |
vojtech | well, our plane is to first finalize the plain BMRM using the SOL intefreace of shogun | 13:17 |
vojtech | I don't know if the interface is finsihed? | 13:17 |
@sonney2k | vojtech, well it is merged but it will only be finished when your BMRM is merged in there | 13:18 |
vojtech | once we have the template for SOL optimizer we will implement our bundle method in C | 13:18 |
vojtech | currently we have Matlab version | 13:18 |
@sonney2k | vojtech, uricamic: so I suggest to have some sub-meeting with n4nd0, nicococo after this session | 13:18 |
@sonney2k | to get it done | 13:18 |
vojtech | is there any algorithm using the SOL interface ? | 13:19 |
vojtech | sonney2k: ok, we can discuss the SOL intefrace in a subgroup | 13:19 |
@sonney2k | vojtech, n4nd0 wrote a summary here https://iglesiashogun.wordpress.com/ | 13:19 |
n4nd0 | vojtech, uricamic: I have started a basic algorithm for SO training, is not yet finished though | 13:19 |
vojtech | I read the pdf he sent | 13:19 |
vojtech | I have queastions but maybe it is better for the SOL people | 13:20 |
vojtech | only | 13:20 |
uricamic | n4nd0: I have read the blog, nice work | 13:20 |
@sonney2k | vojtech, n4nd0, uricamic, alexlistensagain I guess you all are affected and so should continue the discussion in a subgroup (after the meeting?) | 13:20 |
n4nd0 | uricamic: thank you! let me know about your comments and thoughts | 13:20 |
blackburn | #shogun-sol | 13:21 |
blackburn | :D | 13:21 |
@sonney2k | n4nd0, nicococo: so please give us an update :) | 13:21 |
n4nd0 | and that ^ is to all in general, it may be better to introduce changes asap | 13:21 |
n4nd0 | all right | 13:21 |
n4nd0 | so we design the interface for SO learning in shogun | 13:21 |
alexlistensagain | I wanted to discusswith viking where we maybe would like changes in n4ndos framework | 13:22 |
alexlistensagain | will do that today or Tue | 13:22 |
n4nd0 | in this post is more or less explained http://iglesiashogun.wordpress.com/2012/05/22/first-weekly-report-gsoc-2012/ | 13:22 |
n4nd0 | and I think this diagram summarizes it quickly http://iglesiashogun.files.wordpress.com/2012/05/test.pdf | 13:22 |
@sonney2k | I guess that is sufficient | 13:22 |
@sonney2k | for the general public | 13:22 |
alexlistensagain | thx! n4ndo | 13:22 |
@sonney2k | more in the subgroup :) | 13:22 |
@sonney2k | great work btw! | 13:22 |
n4nd0 | and another thing I would like to say | 13:22 |
n4nd0 | yesterday I started to interface mosek from shogun | 13:23 |
@sonney2k | puffin444, oliver - would you please continue after n4nd0 is done? | 13:23 |
n4nd0 | since we need to account for some constraints in the QP | 13:23 |
@sonney2k | n4nd0, ok - thanks | 13:23 |
n4nd0 | I am quite unexperienced with mosek, so if any of you have used it before - specially C interface | 13:23 |
n4nd0 | it will be great to get some suggestions | 13:24 |
@sonney2k | n4nd0, just ask later when you have questions, I did some stuff with cplex' C interface so it should be similar | 13:24 |
@sonney2k | puffin444, please continue | 13:25 |
@sonney2k | oliver, ... | 13:25 |
@sonney2k | ^] | 13:25 |
n4nd0 | sonney2k: all right, thanks | 13:25 |
puffin444 | I have the first weekly milestone tested and ready to be submitted as a patch | 13:25 |
puffin444 | ProductKernel | 13:25 |
alexlistensagain | you use productor geometric average? | 13:26 |
ckwidmer | n4nd0, I guess they are using the python interface, but cvxopt has mosek bindings as far as I know, possibly there is something to learn there | 13:26 |
@sonney2k | puffin444, ahh btw, please make use of CRegressionLabels in your GP code - we now have confidences in there too so please put them there too | 13:26 |
puffin444 | will do | 13:26 |
@sonney2k | puffin444, like a combined kernel but with products K(x,y)* K'(x,y) ? | 13:26 |
puffin444 | Yes elementwise. It's based off of one in the GPML MATLAB toolbox | 13:27 |
alexlistensagain | well product or geometric average -> this is a question about kernel width treatment | 13:27 |
@sonney2k | puffin444, ok | 13:27 |
@sonney2k | puffin444, care to update us about your talk? | 13:27 |
puffin444 | I will be using GPML to make sure my code is accurate | 13:28 |
puffin444 | Yes. I took most of the day Wednesday to travel to Google's offices in Chicago | 13:28 |
puffin444 | I was invited to do a lightning talk, lasting only 5 mins (4 slide limit) | 13:28 |
@sonney2k | pluskid, cheng maybe you can prepare to be next after puffin already now. | 13:29 |
puffin444 | I was able to present the general idea of GP's and how cool shogun was to the audience :) | 13:29 |
@sonney2k | heh | 13:29 |
@sonney2k | oliver, do you have any additional comments? | 13:30 |
@sonney2k | looks like not | 13:30 |
@sonney2k | then thanks puffin444 | 13:30 |
oliver | Great that the product kernel works.... | 13:30 |
oliver | (rorry slow) | 13:30 |
@sonney2k | oliver, anything else? | 13:31 |
oliver | I think it would be good to discuss the framework changes regarding parameter optimization at some point. | 13:31 |
@sonney2k | heiko, that's with you I think | 13:31 |
oliver | That does a bit deeper into the shogun internals. | 13:31 |
heiko | ok | 13:31 |
heiko | I started on a framework for statistical tests | 13:31 |
oliver | Yes, would be great if Heiko has some time to sort this out. There were a couple of emails but it needs a real plan. | 13:31 |
@sonney2k | oliver, pluskid, heiko maybe you schedule some subgroup meeeint | 13:31 |
heiko | ah sorry | 13:32 |
@sonney2k | meeting | 13:32 |
@sonney2k | like the SO guys | 13:32 |
heiko | yes | 13:32 |
nicococo | hi all, (back again ..) | 13:32 |
@sonney2k | nicococo, we were only saying bad things about you | 13:32 |
oliver | that'd be great. | 13:32 |
nicococo | wonderful | 13:32 |
@sonney2k | oliver, heiko, puffin444 maybe directly after the meeting? | 13:32 |
heiko | I am free afterwards but lets talk about this later | 13:32 |
@sonney2k | I think we should continue now | 13:32 |
pluskid | ok | 13:32 |
puffin444 | I think that's what we are doing | 13:33 |
@sonney2k | pluskid, cheng yes | 13:33 |
pluskid | my turn? | 13:33 |
@sonney2k | yes | 13:33 |
pluskid | I have typed a lot :D | 13:33 |
pluskid | I have almost finished the ecoclib porting: one encoder and several decoder left, but I don't plan to implement all of them currently because I think too much (similar) choices might confuse users. I'm currently doing a survey for other MC algorithms. I will post a preliminary draft at the weekly report. | 13:33 |
pluskid | for next, I might move on shareboost or one tree style algorithm from JL | 13:33 |
blackburn | one idea | 13:34 |
blackburn | may be some tree? | 13:34 |
@sonney2k | excellent work... | 13:34 |
pluskid | I'll discuss with Cheng about that next week | 13:34 |
pluskid | I sent an email to the authors of ShareBoost to ask reference code | 13:34 |
@sonney2k | cheng, shareboost was a nips paper last year by shai-shaleve schwhartz | 13:34 |
cheng | I must admit that the credit all goes only to pluskid. I'm sleeping. | 13:34 |
blackburn | TreeMulticlassMachine? | 13:35 |
pluskid | blackburn: Yeah, that's being planed | 13:35 |
@sonney2k | I talked to him and they used it in some glasses for blind people to online do ocr | 13:35 |
pluskid | cheng: you suggested a lot of interesting reading for me | 13:35 |
pluskid | yeah, ShareBoost and many of JL's MC algorithms are aimed at large scale problems | 13:36 |
pluskid | with many many number of classes | 13:36 |
@sonney2k | I think that is where we need to improve - so any algorithm for that would be great | 13:36 |
@sonney2k | almost SO though | 13:36 |
cheng | There is a strange tradeoff, because then there are too few examples per class | 13:37 |
@sonney2k | ok, any final comments | 13:37 |
cheng | With SO learning, the classes are somehow similar. | 13:37 |
@sonney2k | cheng, pluskid? | 13:37 |
cheng | I'm done | 13:37 |
pluskid | cheng: yes, I saw in one of JL's paper that he is using 10^7 examples and 10^6 classes, then there will be only 10 examples in each class, it looks strange for me | 13:37 |
blackburn | while there are a few experts here | 13:37 |
@sonney2k | pluskid, and thanks again for you help in the label / sgvector discussions/transistions. you spent quite some time helping | 13:38 |
blackburn | tell me - is it worth to use some custom distance delta(y_i,y_j) between classes in typical multiclass? | 13:38 |
pluskid | sonney2k: that's because you are soooooooo pushing :D haha | 13:38 |
@sonney2k | hehe | 13:38 |
blackburn | cheng: ^ ? | 13:38 |
blackburn | sonney2k: ? | 13:38 |
alexlistensagain | I have seen some work at ICCV2011 on imagent and large scale learning | 13:38 |
alexlistensagain | on imagenet data | 13:39 |
cheng | blackburn: I think generally it is believed that yes it should help to have custom delta | 13:39 |
pluskid | blackburn: don't know, since we have no structures in the Y space, don't have good knowledge to define a distance other than 0-1 | 13:39 |
@sonney2k | so next one then blackburn and ckwidmer ! | 13:39 |
ckwidmer | well, blackburn has spent some time on fixing up shogun internals and is now working on finishing up Multiclass Domain Adaptation SVM, and will then move back towards merging/implementing stuff from the SLEP package. Also, we'll be merging in some MTL formulation developed with Marius Kloft and nicoco... co. | 13:39 |
cheng | But nobody has shown it to be helpful yet. | 13:39 |
cheng | ckwidmer has some neat paper where they learn similarity between multitask, almost multiclass | 13:39 |
ckwidmer | blackburn, feel free to give more background | 13:40 |
blackburn | hmm | 13:40 |
blackburn | ok | 13:40 |
alexlistensagain | we have also a paper where we learn similarities for multi-task | 13:40 |
@sonney2k | blackburn, besides your usual vodka consumption please | 13:40 |
alexlistensagain | and we won a challenge with it :P | 13:40 |
blackburn | sonney2k: yes this week I spent almost all gsoc money for vodka | 13:40 |
blackburn | about SLEP - we were pretty confused with the way it organized | 13:41 |
@sonney2k | molotov cocktails D: | 13:41 |
blackburn | but currently it is less or more clear | 13:41 |
vojtech | pluskid: imagine you want to classify people according to age categories, then there is clear structure in Y and 0/1 loss is not good | 13:41 |
ckwidmer | alexlistensagain, which paper is that? | 13:41 |
alexlistensagain | Samek et al CAIP2011 :) | 13:41 |
pluskid | vojtech: ok, that makes sense | 13:41 |
alexlistensagain | its uncionventional | 13:41 |
cheng | ckwidmer, blackburn, alexlistenagain: we've got an ICML 2011 paper where we learn the output kernel. Theoretically nice, practically so so | 13:42 |
blackburn | btw is there anybody interested in inverse covariance estimation? | 13:42 |
alexlistensagain | a general delta loss makes sense whenever one wants to penalize one loss over another | 13:42 |
blackburn | slep has some cool routine for that | 13:42 |
@sonney2k | blackburn, ckwidmer I guess that's all from you right? | 13:42 |
ckwidmer | alexlistensagain, cool I'll check it out | 13:42 |
blackburn | yes crappy all | 13:42 |
blackburn | :D | 13:42 |
@sonney2k | So then we should decide when we meet next | 13:43 |
@sonney2k | how about 26.6. 11 UTC again? | 13:43 |
ckwidmer | cheng, the one with invexity, right? | 13:43 |
blackburn | I am ok with any date | 13:43 |
n4nd0 | sonney2k: it is ok for me | 13:43 |
blackburn | except 20 and 14 | 13:43 |
cheng | not ideal, but ok | 13:43 |
uricamic | sonney2k: ok for me | 13:43 |
cheng | ckwidmer: yes invexity | 13:43 |
nicococo | its ok | 13:44 |
pluskid | ok | 13:44 |
ckwidmer | fine for me | 13:44 |
heiko | ok | 13:44 |
alexlistensagain | ok by today | 13:44 |
vojtech | 26.6. is ok | 13:44 |
@sonney2k | OK 26th it is then | 13:44 |
puffin444 | June 26th? | 13:44 |
@sonney2k | pretty many ok's | 13:44 |
@sonney2k | yes | 13:44 |
puffin444 | Seems okay to me | 13:44 |
oliver | fine, yes. | 13:44 |
@sonney2k | at that next meeting you should have 1st half of your work done | 13:44 |
@sonney2k | it is ~1 month till then so plenty time if you work full time | 13:45 |
@sonney2k | ahh and btw an update on gsomix's work | 13:45 |
@sonney2k | he removed CArray* crap from shogun | 13:45 |
@sonney2k | (that was only used in structure/) | 13:45 |
@sonney2k | and replaced it with dynarray | 13:45 |
@sonney2k | introduced CSet/CMap | 13:45 |
@sonney2k | and did lots of cleanups | 13:45 |
@sonney2k | and will now do something pretty cool if we get it to work: | 13:46 |
@sonney2k | director classes | 13:46 |
@sonney2k | i.e., one will be able to overload a shogun method from python | 13:46 |
@sonney2k | with a python function | 13:46 |
heiko | that is awesome! | 13:46 |
@sonney2k | so for prototyping that is crazily cool | 13:46 |
@sonney2k | (think of trying out a new kernel...) | 13:46 |
@sonney2k | anyway that's it form us | 13:47 |
@sonney2k | from us | 13:47 |
@sonney2k | blackburn, anything on your list? | 13:47 |
pluskid | haha, maybe some day shogun is compatible with scikit.learn | 13:47 |
blackburn | no I don't think I have anything | 13:47 |
@sonney2k | or any general questions? | 13:47 |
pluskid | any guidelines/conventions for writing the weekly summary? | 13:47 |
@sonney2k | pluskid, yeah we can use all scikits stuff too :) | 13:48 |
@sonney2k | pluskid, no just *short* | 13:48 |
@sonney2k | I prefer you all program on sth not writing reports | 13:48 |
pluskid | sonney2k: haha, then I'll write "see the IRC log" | 13:48 |
cheng | my 2 cents: weekly reports are more for those of us (like me) who don't hang out too much on IRC... | 13:49 |
@sonney2k | pluskid, seems like I should push you into a different direction :) | 13:49 |
cheng | and still I think they should be short. | 13:49 |
@sonney2k | which reminds me | 13:49 |
@sonney2k | mentors, if possible just idle in #shogun too | 13:49 |
@sonney2k | students rarely ask you directly | 13:49 |
pluskid | cheng: I see :) | 13:49 |
@sonney2k | promised :) | 13:49 |
blackburn | 'sorry I have played world of warcraft whole week' | 13:49 |
@sonney2k | blackburn, 'world of battleship shogun labels' | 13:50 |
blackburn | that would be awesome weekly report | 13:50 |
alexlistensagain | @blackburn: you cannot, you spent all your money on vodka | 13:50 |
pluskid | blackburn: shouldn't be Diablo III? | 13:50 |
eric___ | sonney2k: yes assertion : LibSVM model->nr_class=9 while num_classes=10 | 13:50 |
@sonney2k | alright then if there is nothing more | 13:50 |
@sonney2k | case dismissed | 13:50 |
@sonney2k | thank you all for attending | 13:50 |
ckwidmer | thanks, sonney2k | 13:50 |
n4nd0 | sonney2k: thanks for organizing! | 13:50 |
eric___ | sonney2k: shogun::ShogunException | 13:50 |
cheng | Night! | 13:51 |
alexlistensagain | thanks for organizing Soeren | 13:51 |
n4nd0 | it's quite constructive to have these meetings | 13:51 |
blackburn | alexlistensagain: when can I ask you for some advice? | 13:51 |
blackburn | :) | 13:51 |
alexlistensagain | I hve to go in two hours | 13:51 |
alexlistensagain | otherwise I have time | 13:51 |
blackburn | hmm lets switch to /query then? | 13:51 |
oliver | puffin444, heiko: when shall we talk? | 13:51 |
puffin444 | can we talk now? | 13:52 |
alexlistensagain | we have to see when the SO subgroup will talk | 13:52 |
@sonney2k | eric___, well you call it wrong. but before things didn't get checked: please do obj->io->enable_file_and_line() run the script again | 13:52 |
-!- heiko1 [~heiko@host86-180-47-5.range86-180.btcentralplus.com] has joined #shogun | 13:52 | |
@sonney2k | eric___, and you will see the line number | 13:52 |
@sonney2k | heiko1, ^ puffin444, heiko: when shall we talk? | 13:52 |
oliver | Unfortunately I have a meeting now but can be back in ~1h. | 13:52 |
vojtech | alexlistensagain: yes, we should agree on time for So group | 13:52 |
heiko1 | i could be there in 1h | 13:52 |
alexlistensagain | so blackburn, just ask me now | 13:53 |
-!- heiko [~heiko@host86-185-113-34.range86-185.btcentralplus.com] has quit [Ping timeout: 245 seconds] | 13:53 | |
alexlistensagain | if you like | 13:53 |
@sonney2k | eric___, what are your labels 0...<nr_classes-1> I hope? | 13:53 |
heiko1 | say at 2? | 13:53 |
n4nd0 | vojtech: hey! so you asked before something about structure spaces types | 13:53 |
blackburn | alexlistensagain: hm ok I am working on http://benchmark.ini.rub.de/?section=gtsrb&subsection=news for my bachelor's thesis | 13:53 |
puffin444 | So 1 UTC? | 13:53 |
eric___ | sonney2k: yes, labels 0..9 | 13:53 |
oliver | heiko, puffin444: ok - will be back in1h. I am flexible rest of the day... | 13:53 |
puffin444 | 13 UTC sry | 13:53 |
oliver | yes 1 UTC. | 13:53 |
oliver | great, cu then | 13:53 |
alexlistensagain | ahh, that broken challenge :) | 13:54 |
blackburn | alexlistensagain: why broken? | 13:54 |
vojtech | n4nd0: I need to ready the updated SO document first and then I'd like to ask questions | 13:54 |
alexlistensagain | because people can submit endlessly many results and can see their and others results immediately | 13:54 |
n4nd0 | vojtech: ok | 13:54 |
alexlistensagain | then you do not need to do model selection | 13:55 |
blackburn | alexlistensagain: I stucked at 97.1% accuracy - need to improve features | 13:55 |
alexlistensagain | thats why I consider it broken | 13:55 |
vojtech | n4nd0: are you the person to ask ? | 13:55 |
alexlistensagain | ok | 13:55 |
n4nd0 | vojtech: I am doing generic SO project so probably yes - I hope I'll be able to answer :) | 13:55 |
blackburn | alexlistensagain: what do you think is good normalization before computing HOGs? | 13:56 |
eric___ | sonney2k: error from MulticlassLibSVM.cpp line 102 | 13:56 |
n4nd0 | uricamic: do you think the structure we have designed for SO fits well for the bundle algorithms you will implement? | 13:56 |
vojtech | n4nd0: 1st quick question: why do you need the generic constraints on w, like A'*w = b ? | 13:56 |
n4nd0 | vojtech: ok, so it is nicococo who wrote that pdf | 13:57 |
-!- heiko1 [~heiko@host86-180-47-5.range86-180.btcentralplus.com] has quit [Ping timeout: 245 seconds] | 13:57 | |
vojtech | n4nd0: do you want to be as generic as possible or is it motivated by some practical SO tasks? | 13:57 |
-!- puffin444 [62e3926e@gateway/web/freenode/ip.98.227.146.110] has quit [Quit: Page closed] | 13:57 | |
n4nd0 | vojtech: and yesterday he told me I should forget about all those constraints | 13:57 |
eric___ | sonney2k: the call is from mc Xval function | 13:57 |
vojtech | n4nd0: :) | 13:57 |
nicococo | vojtech: we are actually using those constraints for gene finding application | 13:57 |
n4nd0 | vojtech: and just consider Aw <= b and lb <= w <= ub | 13:57 |
-!- heiko [~heiko@host86-176-176-185.range86-176.btcentralplus.com] has joined #shogun | 13:58 | |
heiko | sorry, connection hazzle :( | 13:58 |
vojtech | nicococo: we also use Aw ,= b constraints | 13:58 |
vojtech | aw<= b | 13:58 |
uricamic | n4nd0: yep, I think it is ok, the only problem could be the regularization-matrix C, I have now the implementation just for a single regularizer | 13:58 |
vojtech | nicococo: but is it possible that some SO solvers will optimize a special instance of the generic SO problem you defined? | 13:59 |
nicococo | vojtech: i think it might not be necessary that every SO solver should be able to handle ALL constraints.. it is fine, if they have their restrictions | 13:59 |
vojtech | nicococo: yes, this is reasonable because exploiting a special (simpler) structure of the SO problem can significantly decree time | 14:00 |
-!- cheng [73406f11@gateway/web/freenode/ip.115.64.111.17] has quit [Quit: Page closed] | 14:01 | |
nicococo | vojtech: yepp, the solver could give a hint (msg) if it is not compatible with the problem defined. | 14:02 |
vojtech | nicococo: the PDF document also mentioned slack-rescaling loss approximation. is it still the plane? if yes I wonder for which SO tasks you can optimize it? | 14:02 |
nicococo | vojtech: we have a first sketch of slack rescaling for hm-svms, also there is an icml paper 2009 | 14:03 |
nicococo | but i don't know if it boosts performance :( | 14:04 |
vojtech | nicococo: I experimented with the PosLearn loss function (ICML 2009) and it does not work for me at all .... | 14:04 |
nicococo | vojtech: i think the bound is really loose | 14:05 |
nicococo | ..on page 3 Eqn. (2) | 14:06 |
vojtech | nicococo: I compared the PosLearn loss with margin-rescaling loss on synthetic data and margin-loss clearly wins | 14:06 |
nicococo | uhh.. | 14:06 |
vojtech | nicococo: By PosLearn loss I mean the problem on page 4 | 14:07 |
nicococo | vojtech: so all constraints get a slack, right? | 14:07 |
vojtech | nicococo: I didn't experiment with the variational approximation of slack-rescaling | 14:07 |
vojtech | yes | 14:08 |
nicococo | so its a bit like max-margin markov networks | 14:08 |
vojtech | I don't think, max-margin MN use standard margin-loss | 14:08 |
vojtech | if you mean papers by Taskar | 14:08 |
vojtech | they just replace the augmented max problem with LP relaxation | 14:09 |
vojtech | you mentioned that ye tried slcak-rescaling for HMMSVM, any results ? | 14:10 |
nicococo | vojtech: right :) | 14:10 |
nicococo | no so far we have no results :( but at least we know there is a linear time viterbi-style approximation possible.. | 14:11 |
eric___ | sonney2k: is there a way to print subsets indices (in xval) without recompiling shogun ? | 14:11 |
sonne|work | eric___: heiko is the xval expert (he wrote the code) | 14:11 |
vojtech | nicococo: do you mean the variatinal approximation from the icml2009 paper or something new ? | 14:12 |
nicococo | vojtech: as soon as i have results i can tell you ... | 14:12 |
nicococo | no it is something different | 14:12 |
heiko | eric___ what do you want to do? | 14:12 |
heiko | there is no method for that, so recompiling would be necessary | 14:12 |
nicococo | vojtech: we are solving the argmax delta(1-dPsi) directly | 14:13 |
vojtech | nicococo: it is very interesting. I'll be happy to learn about it ... | 14:13 |
sonne|work | I eric___but I guess in this case you have fewer classes in this data split than in your labels | 14:13 |
heiko | I could add a debug message if you want, eric___ | 14:13 |
sonne|work | so no example from class 8 or 9 | 14:13 |
nicococo | vojtech: i will have results 1 or 2 weeks after nips deadline (for toy data at least, and compared against margin rescaling) | 14:14 |
eric___ | heiko: yes it would be nice | 14:14 |
heiko | k | 14:14 |
heiko | what exctly do you want? | 14:15 |
eric___ | heiko: I got some strange results using multiclass xvalid | 14:15 |
heiko | training/testing indices in x-val? | 14:15 |
eric___ | heiko: yes | 14:15 |
heiko | ok | 14:15 |
heiko | will do, give me a few minutes | 14:15 |
vojtech | nicococo: thanks | 14:15 |
eric___ | heiko: I got always "XVal: [0.100000,0.100000] with alpha=0.050000, mean=0.100000" in mc xval | 14:15 |
nicococo | vojtech: a pleasure :) | 14:15 |
heiko | that means all results were equal | 14:16 |
sonne|work | pluskid: actually weird how can eric___'s assert even be triggered | 14:16 |
nicococo | i have to go to lunch now.. i am starving.. bye | 14:16 |
sonne|work | pluskid: in bool CMulticlassLibSVM::train_machine(CFeatures* data) | 14:16 |
sonne|work | line 40 | 14:16 |
n4nd0 | nicococo: bye, enjoy your meal! | 14:16 |
sonne|work | you get the number of classes | 14:16 |
eric___ | heiko: I cannot send you the features, too big, but when I split the data manually I got mean > 0.7 | 14:16 |
sonne|work | which is 9 | 14:16 |
sonne|work | and labels have 10 classes | 14:16 |
heiko | eric___, or than there is a problem | 14:16 |
-!- nicococo [~nico@lacedcoffee.ml.tu-berlin.de] has left #shogun [] | 14:17 | |
heiko | might also be because of changes in Labels system | 14:17 |
pluskid | strange | 14:17 |
heiko | I had some problems with that a few days ago | 14:17 |
sonne|work | eric___: what are your labels? | 14:17 |
sonne|work | 0...9 ? | 14:17 |
sonne|work | or something else in there? | 14:17 |
eric___ | sonne|work: yes | 14:17 |
eric___ | sonne|work: digits classes | 14:17 |
pluskid | sonney2k: "labels have 10 classes" <-- is this returned by labels->get_num_classes() ? | 14:18 |
pluskid | the number in the strategy is returned by such a call | 14:19 |
pluskid | in MulticlassMachine.cpp line 60 | 14:19 |
sonne|work | pluskid: he has the error (01:50:27 PM) eric___: sonney2k: yes assertion : LibSVM model->nr_class=9 while num_classes=10 | 14:19 |
sonne|work | pluskid: in multiclasslibsvm.cpp line 102 | 14:20 |
pluskid | so where is model->nr_class calculated? | 14:20 |
eric___ | sonne|work: pluskid: m LibSVM model->nr_class=9 while num_classes=10 | 14:20 |
sonne|work | pluskid: I guess in libsvm... | 14:21 |
sonne|work | maybe one class is missing | 14:21 |
sonne|work | so he would have labels 0...8,9 | 14:22 |
sonne|work | err 0...4,6,...9 | 14:22 |
sonne|work | 5 missing for example | 14:22 |
pluskid | why missing? | 14:22 |
-!- vojtech [9320543b@gateway/web/freenode/ip.147.32.84.59] has quit [Quit: Page closed] | 14:25 | |
sonne|work | pluskid: maybe 5 is not in the labels and libsvm checks the actual number of classes ocurring... | 14:25 |
eric___ | pluskid: sonne|work: its happen when I call cross->evaluate(). In the same time I run manual split of the base train/test and I don't have any error. Mb it helps ? | 14:25 |
pluskid | sonne|work: but doesn't CLabels->get_num_class check that? it use CMath::unique IIRC | 14:26 |
pluskid | oops, something related to cross validation component? | 14:27 |
eric___ | I checked my labels, I have 50 samples; 5 of each classes and my xval is launch with n_folds = 5 | 14:28 |
heiko | sonne|work, is there any way to output a vector in only debug mode? | 14:31 |
heiko | because calling SG_DEBUG multiple times always has this ugly [DEBUG] before every element | 14:32 |
sonne|work | heiko: well you can call get_loglevel() | 14:32 |
heiko | perfect :) | 14:32 |
sonne|work | and then only output sth if io->get_loglevel() <= MSG_DEBUG | 14:33 |
sonne|work | pluskid: true then num_classes shouldn't be 10 | 14:35 |
sonne|work | hmmhh | 14:35 |
CIA-113 | shogun: Heiko Strathmann master * rf377998 / (2 files in 2 dirs): added debug messages for x-val - http://git.io/Yt7J5A | 14:38 |
CIA-113 | shogun: Heiko Strathmann master * ra4fd65a / (2 files in 2 dirs): Merge pull request #553 from karlnapf/master - http://git.io/qEEoiw | 14:38 |
heiko | eric___ the indices are now printed | 14:38 |
heiko | I will grab some food now, be back at 2 | 14:39 |
eric___ | heiko: thx | 14:42 |
eric___ | heiko: debug will be easier, I let you know | 14:42 |
-!- pluskid [~pluskid@111.120.53.76] has quit [Ping timeout: 256 seconds] | 14:44 | |
-!- pluskid [~pluskid@li400-235.members.linode.com] has joined #shogun | 14:44 | |
-!- puffin444 [62e3926e@gateway/web/freenode/ip.98.227.146.110] has joined #shogun | 14:51 | |
-!- n4nd0 [~nando@s83-179-44-135.cust.tele2.se] has quit [Quit: leaving] | 14:52 | |
heiko | pluskid, are you there? | 15:01 |
puffin444 | Hi Heiko. Is Oliver here yet? | 15:04 |
heiko | hi puffin444, not yet I think | 15:04 |
heiko | ehm, I get all these nicknames mixed up | 15:05 |
heiko | mixed you up with pluskid :) | 15:05 |
puffin444 | I confused you with blackburn for couple of days :) | 15:06 |
heiko | lol :) | 15:06 |
heiko | lets way a few minutes | 15:06 |
heiko | oliver, are you around? | 15:06 |
eric___ | heiko: how do I print subsets indices ? | 15:07 |
heiko | eric___, sg_io->set_loglevel(MSG_DEBUG); | 15:07 |
heiko | after shogun init | 15:07 |
heiko | there will be some other stuff poping up too | 15:08 |
eric___ | heiko: tx | 15:08 |
heiko | puffin444, what should be discuss? | 15:10 |
puffin444 | Do you have a copy of the model selection diagrams I emailed some days ago? | 15:11 |
heiko | Wait I will get it | 15:12 |
heiko | k | 15:13 |
puffin444 | so what I had in mind was abstracting away crossvalidation into a machineevaluation class that could do all kinds of different evaluations | 15:14 |
heiko | yes I lked that | 15:14 |
oliver | hi all | 15:14 |
oliver | back now | 15:14 |
oliver | sorry; took a bit longer. | 15:14 |
heiko | oliver, no worries, we just started | 15:14 |
puffin444 | GradientOptimization inherits from machineevaluation adding an ability to obtain a gradient, as well as a differentiable function to be optimized | 15:14 |
puffin444 | hi oliver | 15:14 |
puffin444 | oliver, do you have a copy of the uml diagram I emailed some time back? | 15:15 |
pluskid | I made a list of nickname -> name -> project mapping, but lost that file when I re-install my system... too bad | 15:15 |
heiko | puffin444, why do you need this extra ability to obtain the gradient? | 15:16 |
heiko | this should already be doable via the "evaluate" method | 15:16 |
oliver | yes, have this here. | 15:16 |
heiko | which should be renamed | 15:16 |
puffin444 | I thought that only specific forms of evaluation use a gradient, so I put it in a specialized class | 15:16 |
puffin444 | Perhaps it makes more sense to put it in MachineEvaluation? | 15:16 |
heiko | what about a method | 15:17 |
sonne|work | pluskid: who's real name do you miss? | 15:17 |
heiko | mmh | 15:17 |
-!- uricamic [~uricamic@2001:718:2:1634:5c56:9bc4:b0a1:c02b] has quit [Quit: Leaving.] | 15:17 | |
puffin444 | A method in MachineEvaluation that returns the gradient? | 15:17 |
heiko | I mean GradienResult already inherits from evaluation result | 15:17 |
heiko | so you could just return it like that | 15:17 |
pluskid | sonne|work: I'm not good at remember names, but I guess I can rebuild the list from the self-intro emails if necessary | 15:17 |
heiko | via the evaluation method | 15:18 |
heiko | and the GradientOptimisation class would then overload that method | 15:19 |
heiko | and return the gradient | 15:19 |
puffin444 | Okay, so the gradient is stored in GradientResult | 15:19 |
puffin444 | I see where you are coming from. | 15:20 |
heiko | I would like to generalise as much as possible | 15:20 |
puffin444 | Does it make sense for GradientOptimization to have its own member called differentiablefunction | 15:21 |
heiko | This evaluation class just returns the quantity that is used for optimisation | 15:21 |
puffin444 | okay | 15:21 |
heiko | yes, I think so | 15:21 |
oliver | I am missing a bit the function that is evaluated. | 15:21 |
oliver | Where does this come in? | 15:21 |
puffin444 | That is the differentiable function class | 15:21 |
oliver | There is no need for GP methods to focus on marginal likelihood exclusively. | 15:21 |
puffin444 | quantity is what is to be optimized, and gradient represents the gradient of the function that returns the quantity | 15:22 |
oliver | yes but why not have "objectiveFunction" which is either "Function" or "DifferentiableFunction" | 15:22 |
oliver | ? | 15:22 |
oliver | THere is no need to optimize the marginal likelihood using gradients per se (there are some good reasons to use grids as well in some cases). | 15:22 |
puffin444 | So a function that may or may not have gradients? | 15:23 |
oliver | It is merely an option as you can evaluate the gradient... | 15:23 |
oliver | yes] | 15:23 |
heiko | yes, but that is than not handled via the Gradient class | 15:23 |
heiko | but for example by gridsearch | 15:23 |
oliver | I think that is the standard interface an optimizer would use (perhaps worth looking at NLOPT for typical interfaces used in optimization). | 15:23 |
heiko | and for all function we would need a specialisation of MachineEvaluation | 15:23 |
oliver | I see two levels. | 15:23 |
oliver | a) the objective which can be marginal likelihood or classification performance or similar. | 15:24 |
heiko | eg. for the marginal likelihood we have another class that inherits from MachineEval | 15:24 |
heiko | and then this class returns the marginal likelihood via the evaluate method | 15:24 |
heiko | another example is x-val | 15:24 |
oliver | b) The optimization which can be grid based or by means of gradient optimization | 15:24 |
heiko | oliver yes | 15:24 |
heiko | split the objective (gradient, xval, objective) and its optimisation | 15:25 |
oliver | yes, agree | 15:25 |
heiko | Thats why I think all evaluations (whether they are gradients, objective values, or x-val estimates) should be accessible via the MachineEvaluation:evaluate method | 15:25 |
eric___ | heiko: can I send your somewhere the debug mess of my xval, you may help me then find the problem ? | 15:26 |
puffin444 | Okay, so theres XValidation, GradientOptimzaition, and ObjectiveOptimization that all inherit from MachineEvaluation? | 15:26 |
heiko | eric___, what about rather reproducing the problem in a minimal example? It takes ages for me to get into a large debug mess :) | 15:26 |
heiko | puffin444, yes | 15:27 |
heiko | but remove the optimisation | 15:27 |
heiko | that is happening in another class | 15:27 |
puffin444 | Okay. | 15:27 |
heiko | currently only evaluating the criteria for optimisation | 15:27 |
heiko | I am also thinking of a more informative name than MachineEvaluation ... mmh? | 15:28 |
heiko | actually its fine :) | 15:28 |
oliver | yeah... | 15:29 |
oliver | one deatil: for LML it make sense to optimize ont he training set also. | 15:29 |
oliver | I.e. you want a cross validation class that con just carry out a grid search on the training set. | 15:29 |
puffin444 | Another question. How will Gradient/Objective get access to the function to be optmized? | 15:30 |
oliver | It's then barely anything else a grid based global optimizer but would be good to have this functinality as gradinet-based stuff runs into | 15:30 |
oliver | local optima quite often. | 15:30 |
oliver | Alternatively one could stick in a global optimizer. | 15:30 |
heiko | oliver, yes, I think there should be a flag in the GridSearchModelSelection class | 15:30 |
oliver | puffin444: What optimizer are you planing to use? | 15:30 |
eric___ | heiko: in the same xval run, printed training set indices and testing set indices are the same | 15:30 |
puffin444 | Say I want to optimze the likihood using gradient search. | 15:31 |
heiko | eric___, mmmh thats probably multiclass related stuff then, or labels. Could you isolate that in a small program? Then I can check | 15:32 |
puffin444 | Should the function be directly set independently of the machine (which seems weird), or should it somehow extract it from the Machine? | 15:32 |
puffin444 | *the optimizer | 15:32 |
heiko | puffin444 what do you mean? | 15:33 |
heiko | the differentiable function? | 15:33 |
puffin444 | Yest | 15:33 |
puffin444 | Yes | 15:33 |
heiko | good question :) | 15:34 |
puffin444 | Should the machine have some getter/setter, or should the differential function just be set independently of the machine? | 15:34 |
heiko | It clearly depends on the machine's parameters right? | 15:34 |
heiko | what happens if you would add a get_lh_gradient to the GP base class? | 15:35 |
puffin444 | I think so. Theoretically if the differentiable function is independently set, it may not have any relation whatsoever to the machine | 15:35 |
puffin444 | which seems silly | 15:35 |
heiko | indeed | 15:35 |
heiko | but this happens when we generalise in these two layers | 15:36 |
heiko | what about obtaining the function from the machine then? | 15:36 |
heiko | mmmh, but thats difficult with types | 15:36 |
oliver | hm... I am confused still. | 15:36 |
heiko | No I think one will have to do it by hand | 15:36 |
puffin444 | Okay. | 15:37 |
heiko | but if you got a better solution? you are right its weird that one can set completely different functions | 15:37 |
oliver | Sorry, I missed this earlier. Wat exactly is the purpose of machine evaluation? | 15:37 |
oliver | seems like cross validation and gradientOptimization are fully capable of doing the job on their own. | 15:37 |
heiko | to return the value that you are optimising | 15:37 |
heiko | oliver, MachineEvaluation is the generalisation of x-val | 15:38 |
heiko | which does only evaluate the machine | 15:38 |
heiko | or the gradient on its objective | 15:38 |
oliver | But then MachineEvalaution *is* the function to be optimized. | 15:39 |
eric___ | heiko: in the libshogun mcxval example, printed indices for train set 0 and test set 0 are the same. | 15:39 |
heiko | eric___ ok I will have a look little later | 15:40 |
heiko | yes | 15:40 |
oliver | To put the question the other way round: where would the "Gaussian process class" be in this setting? | 15:41 |
heiko | it would be in the machine variable of MachineEvaluation | 15:41 |
heiko | and then either | 15:41 |
eric___ | heiko: ok thx | 15:41 |
heiko | a) x-val class is asked for estimate by grid-search | 15:42 |
heiko | b) gradient evaluation class is asked for gradient by GradientModelSelection | 15:42 |
oliver | ok - sorry; I think I got it now ;) | 15:42 |
heiko | both x-val class and gradient eval class are subclasses of machine evaluation class | 15:42 |
oliver | There is a fundamental difference between a/b, right. | 15:42 |
heiko | oliver, yes :) | 15:43 |
oliver | I actually mean: the marginal likelihood can be directly calculated from the GP class, no "evaluation" needed. | 15:43 |
oliver | Any test set error requires the "Machine Evaluation" to wrap around. | 15:43 |
heiko | For that we could add another subclass of MachineEvaluation that returns an objective function | 15:44 |
oliver | yes | 15:44 |
heiko | for the cases where that can be computed instantly | 15:44 |
heiko | and that may then e.g. be called from grid-search opti | 15:45 |
heiko | thats why I think the name evaluation may be misleading | 15:45 |
oliver | true | 15:45 |
heiko | its more a class to generalise different numerical estimates that have to be max/minimised/0ed | 15:45 |
oliver | How about: MachineEvaluation provides the interfaces for either grid-based or graident based optimization (in principle). | 15:46 |
oliver | The different types of evaluations could then just be sub classes. | 15:46 |
oliver | MachienEvaluationMSE | 15:46 |
oliver | MachineEvaluationLML (log marginal likelihood), etc. | 15:46 |
heiko | We have the model selection class | 15:47 |
heiko | which currently has the x-val for evaluation | 15:47 |
heiko | this could be replaced by the MachineEval class | 15:47 |
heiko | and then subclasses of model selection could call the evaluate method of MachineEval | 15:47 |
heiko | then it would be split a bit more | 15:48 |
heiko | and new methods for model selection would only require adding a new subclass of model selection | 15:48 |
oliver | yes, that sounds nice | 15:48 |
oliver | Perhaps useful to think about the workflow of the user. | 15:49 |
heiko | currently, te grid-search works on x-val. With that system it would work with any subclass of MachineEval | 15:49 |
heiko | same for gradient stuff (not implemented yet) | 15:49 |
heiko | would work on gradient of ML as well as on gradient of some SVM-C gradient | 15:49 |
heiko | and this hybrid case grid/gradient would be another instance of ModelSelection that uses both of the other classes | 15:50 |
heiko | yes, it would be nice to have some eas< methods for everything | 15:50 |
heiko | that do all the technical pluggin classes into each other | 15:50 |
heiko | if you look at the current model selection examples, these are pretty complicated | 15:51 |
oliver | Agree, but it is complicated to be general. | 15:51 |
oliver | Nevertheless we need to start somewhere. | 15:51 |
heiko | I think a good starting point is to do this MachineEvaluation class | 15:51 |
oliver | puffin444: perhaps most useful if you start with the general interface to evaluation marginal likelihoods and gradients, but encapsulating them into a general framework of differentiable functions etc. | 15:52 |
oliver | Those would then be "naively" optimizable by calling a gradient optimizer and can in due course be integrated | 15:52 |
oliver | with the heavy machinery... | 15:52 |
heiko | thats a nice idea | 15:52 |
oliver | I think it's good to keep things simple until the gradients are all working / matching and you get results that are compataible with existing libraries. | 15:53 |
puffin444 | okay. specific to general | 15:53 |
oliver | FOr debuging the full framework seems a pain.... | 15:53 |
oliver | Exactly. Let's make it work first. Simple and pragmatic, but of course having the final solution in mind. | 15:53 |
heiko | oliver, yes thats your part - I currently care about the framework :) | 15:54 |
oliver | sure thing - and that's very much needed! | 15:54 |
oliver | I just know the pain of debugging gradients and figuring out why they don't much what we expect. | 15:54 |
oliver | I think it's key that the basic GP parameter optimization works somehow first. | 15:54 |
heiko | hehe, yes | 15:54 |
heiko | puffin444, you should write a lot of tests for the isolated gradient methods | 15:55 |
heiko | with defined inputs/outputs | 15:55 |
puffin444 | By the way, oliver, do you have any concerns about the diagrams about GPs I made some time bach? | 15:55 |
puffin444 | keiko: will do | 15:55 |
heiko | shogun lacks that, and its so hard to find errors without it | 15:55 |
puffin444 | *back | 15:55 |
blackburn | keiko matsui :D | 15:55 |
puffin444 | :) lol | 15:55 |
oliver | puffin444: I think it looks verygood and matches the previous discussions. Definitely a very good start. | 15:56 |
oliver | Not sure down the line whether the GetAlpha stuff for the inference method is general enough. | 15:56 |
puffin444 | According to the project plan I will be implementing those in the next few weeks first before modelselection | 15:56 |
heiko | blackburn, everybody asks me if my name is Japanese here in London :) but ITS NOT ITS GERMAN | 15:56 |
oliver | From my experience I think it's best if you get something to work first and then you'll see that several plans need to be adapted here andt here. | 15:57 |
blackburn | well they just pronounce it in a wrong way I guess | 15:57 |
oliver | The same will happen to the framework integration of the optimization. | 15:57 |
heiko | oliver, puffin444, yes the modsel stuff is not really important now | 15:57 |
oliver | but still good to have discussed this. | 15:57 |
heiko | needs to be changed all over a few times anyway (my last years experience :) | 15:57 |
puffin444 | I already have for this week the productkernel written and tested. | 15:58 |
puffin444 | I just want to make sure there are no memory leaks before I submit it as a patch | 15:58 |
oliver | Nice! | 15:58 |
oliver | But feel free to submit it early. I have time tomorrow to look at things (weekend is generally good). | 15:59 |
puffin444 | Well it's on the project plan for this week. | 15:59 |
puffin444 | I don't want to get behind! :) | 15:59 |
heiko | oliver, puffin444, thanks for the discussion, I gotta go now (studygroup for exam) | 15:59 |
puffin444 | Thanks for your time, heiko. | 16:00 |
oliver | great, thanks for inputs. | 16:00 |
oliver | and talk soon. | 16:00 |
heiko | take care everybody! | 16:00 |
heiko | bye | 16:00 |
puffin444 | oliver, if you don't mind, I would like to go back to bed. | 16:00 |
puffin444 | bye | 16:00 |
oliver | ok, sure - bye! | 16:00 |
puffin444 | thanks for taking time for this meeting, oliver. | 16:00 |
oliver | Perhaps let's try to talk sometime soon again. | 16:00 |
puffin444 | Absolutely. What time would be good for you? | 16:01 |
oliver | I am around over the weekend... | 16:01 |
oliver | tomorrow is perfect; next week I am travelling quite a bit but certainly can come online here and there. | 16:01 |
puffin444 | What time tomorrow would be good? | 16:01 |
oliver | UTC 8 am-UTC4 pm | 16:02 |
oliver | Just write me an email | 16:02 |
puffin444 | I will. | 16:03 |
puffin444 | See you all later | 16:03 |
oliver | cu | 16:03 |
-!- oliver [c07c1afb@gateway/web/freenode/ip.192.124.26.251] has quit [Quit: Page closed] | 16:03 | |
-!- puffin444 [62e3926e@gateway/web/freenode/ip.98.227.146.110] has quit [Quit: Page closed] | 16:03 | |
-!- alexlistensagain [c25fae8d@gateway/web/freenode/ip.194.95.174.141] has quit [Quit: Page closed] | 16:10 | |
CIA-113 | shogun: Heiko Strathmann master * r07554b5 / src/shogun/evaluation/CrossValidation.cpp : corrected debug message - http://git.io/PHWNRA | 16:12 |
CIA-113 | shogun: Heiko Strathmann master * r43b5b5e / src/shogun/evaluation/CrossValidation.cpp : Merge pull request #554 from karlnapf/master - http://git.io/u6eoPw | 16:12 |
-!- heiko [~heiko@host86-176-176-185.range86-176.btcentralplus.com] has left #shogun [] | 16:28 | |
eric___ | sonne|work: is it possible to enable [DEBUG] messages without [INFO] ones | 16:46 |
CIA-113 | shogun: Soeren Sonnenburg master * r1de9b14 / (32 files): fix java modular examples - http://git.io/8T7fkQ | 17:19 |
blackburn | sonney2k: alex helped me to overwhelm 97.00% curse! :D | 17:25 |
-!- eric___ [2e1fd566@gateway/web/freenode/ip.46.31.213.102] has quit [Quit: Page closed] | 17:31 | |
-!- pluskid [~pluskid@li400-235.members.linode.com] has quit [Quit: Leaving] | 17:31 | |
-!- cwidmer [~chris@HSI-KBW-046-005-237-106.hsi8.kabel-badenwuerttemberg.de] has joined #shogun | 18:16 | |
-!- ckwidmer [~chris@HSI-KBW-046-005-237-106.hsi8.kabel-badenwuerttemberg.de] has quit [Ping timeout: 246 seconds] | 18:25 | |
-!- gsomix [~gsomix@178.45.68.181] has joined #shogun | 18:27 | |
gsomix | hi all | 18:27 |
gsomix | home, sweet home | 18:27 |
gsomix | "@SpaceX #Dragon capsule securely bolted to the Harmony module of the #ISS at 12:02pmET!" | 18:35 |
gsomix | ok, time to work | 18:35 |
-!- cwidmer [~chris@HSI-KBW-046-005-237-106.hsi8.kabel-badenwuerttemberg.de] has quit [Read error: Operation timed out] | 18:51 | |
-!- ckwidmer [~chris@HSI-KBW-046-005-237-106.hsi8.kabel-badenwuerttemberg.de] has joined #shogun | 19:27 | |
@sonney2k | blackburn, to get what? | 19:28 |
blackburn | sonney2k: to get what? | 19:29 |
blackburn | :) | 19:29 |
blackburn | I do not understand | 19:29 |
@sonney2k | 97.00001% | 19:29 |
@sonney2k | ? | 19:29 |
blackburn | hah no | 19:29 |
blackburn | 97.58 | 19:29 |
@sonney2k | heh | 19:29 |
blackburn | sonney2k: btw I have interesting problem here | 19:29 |
blackburn | we lost information about outputs in multiclass machine | 19:30 |
blackburn | so we can't plot OvR ROC currentyl | 19:30 |
@sonney2k | blackburn, ehh? we never had outputs in multiclassmachine | 19:39 |
@sonney2k | just 0,1,... | 19:39 |
blackburn | sonney2k: yes but need to have | 19:40 |
-!- gsomix_ [~gsomix@109.169.139.36] has joined #shogun | 19:42 | |
-!- gsomix [~gsomix@178.45.68.181] has quit [Read error: Connection reset by peer] | 19:42 | |
@sonney2k | blackburn, with new labels you can store confidences | 19:44 |
@sonney2k | so feel free to add the max f(x) .. | 19:44 |
blackburn | yes but one vector = # class confidences | 19:44 |
blackburn | it won't help to plot ROC | 19:44 |
CIA-113 | shogun: Soeren Sonnenburg master * rf96fdfa / (24 files in 2 dirs): fix c# and ruby example to work with new SGVector - http://git.io/u50cgw | 19:47 |
@sonney2k | blackburn, you can always overload the confidence system for multiclass... | 19:47 |
blackburn | he | 19:47 |
@sonney2k | it can be huge though | 19:47 |
@sonney2k | blackburn, while I am away - we need more coverage of our examples in languages other than python | 19:53 |
blackburn | while you are away? | 19:53 |
@sonney2k | well beginning of june | 19:54 |
blackburn | ah | 19:54 |
@sonney2k | it would make a lot of sense to get all the octave_modular etc examples updtaed with what we have in python | 19:54 |
blackburn | should be easy but takes time | 19:54 |
@sonney2k | no difficult task and I think one can do 1-2 langueages per day | 19:55 |
@sonney2k | and I think it is actually partially interesting | 19:55 |
@sonney2k | I mean learnign the basics of 5 $LANG :) | 19:55 |
@sonney2k | man the buildbot is overloaded | 19:56 |
blackburn | 97.67! | 19:56 |
blackburn | :D | 19:56 |
@sonney2k | blackburn, what is the secret ingredient? | 19:56 |
blackburn | C :D | 19:56 |
@sonney2k | you didn't do model selection?!? | 19:57 |
@sonney2k | c'mon!? | 19:57 |
blackburn | sonney2k: I do not have to do model selection - I know correct results | 19:57 |
@sonney2k | ??? | 19:57 |
blackburn | I just fit C for best accuracy :D | 19:57 |
@sonney2k | that is what one calls modelselection | 19:58 |
@sonney2k | gridsearch in shogun :) | 19:58 |
blackburn | no I have test/train features and labels | 19:58 |
@sonney2k | exactly | 19:58 |
blackburn | okay actually I am changing features | 19:59 |
-!- gsomix_ is now known as gsomix | 19:59 | |
@sonney2k | ?? | 20:00 |
@sonney2k | C is not the C of the SVM? | 20:01 |
blackburn | ??! :D | 20:01 |
blackburn | yes I vary over C | 20:01 |
blackburn | but actually I changed features | 20:01 |
blackburn | oook I need 98%! | 20:07 |
blackburn | :D | 20:07 |
CIA-113 | shogun: Sergey Lisitsyn master * reac2e3c / (5 files in 2 dirs): Added rejection strategy handling to multiclass machine - http://git.io/EKOONQ | 21:03 |
-!- gsomix [~gsomix@109.169.139.36] has quit [Quit: Ex-Chat] | 21:16 | |
CIA-113 | shogun: Sergey Lisitsyn master * ra077bd5 / src/shogun/multiclass/RejectionStrategy.h : Fixes for Qtest rejection strategy - http://git.io/yrKL2g | 21:34 |
-!- n4nd0 [~nando@s83-179-44-135.cust.tele2.se] has joined #shogun | 22:11 | |
-!- ckwidmer [~chris@HSI-KBW-046-005-237-106.hsi8.kabel-badenwuerttemberg.de] has quit [Remote host closed the connection] | 22:15 | |
@sonney2k | blackburn, yet another toolbox https://github.com/davidpicard/jkernelmachines/wiki | 22:23 |
@sonney2k | we should probably make sure we have all the kernels they have :) | 22:23 |
blackburn | WTF is adaptative | 22:24 |
blackburn | :D | 22:24 |
blackburn | sonney2k: looks like useless crap at the very first sight | 22:26 |
@sonney2k | yeah indeed | 22:34 |
blackburn | sonney2k: my Q-test idea failed it seems | 22:34 |
blackburn | :D | 22:34 |
@sonney2k | hmmh the buildbot has no chance to catch up it seems... | 22:34 |
@sonney2k | Q-test? | 22:34 |
blackburn | yes I tried to determine rejects according to Q-test for outlier detection | 22:35 |
blackburn | i.e. if score can be considered as outlier | 22:35 |
blackburn | it should not be rejected | 22:35 |
blackburn | else reject | 22:35 |
blackburn | like [3.0,0.1,0.2] - not reject and [0.3,0.1,0.2] - reject | 22:36 |
@sonney2k | ok | 22:45 |
@sonney2k | makes sense | 22:45 |
@sonney2k | anyway | 22:45 |
@sonney2k | tme to slep | 22:45 |
@sonney2k | :D | 22:45 |
@sonney2k | oh swig 2.0.6 release | 22:46 |
@sonney2k | d | 22:46 |
blackburn | sonney2k: slep? | 22:47 |
blackburn | yes right | 22:47 |
blackburn | time to SLEP | 22:47 |
blackburn | :D | 22:47 |
shogun-buildbot | build #560 of java_modular is complete: Success [build successful] Build details are at http://www.shogun-toolbox.org/buildbot/builders/java_modular/builds/560 | 23:06 |
--- Log closed Sat May 26 00:00:41 2012 |
Generated by irclog2html.py 2.10.0 by Marius Gedminas - find it at mg.pov.lt!