--- Log opened Mon Jul 30 00:00:17 2012 | ||
-!- emrecelikten [~emre@176.40.234.211] has left #shogun [] | 00:33 | |
-!- emrecelikten [~emre@176.40.234.211] has joined #shogun | 00:33 | |
-!- needsch [~user@ip-176-198-229-75.unitymediagroup.de] has quit [Ping timeout: 246 seconds] | 00:41 | |
-!- emrecelikten [~emre@176.40.234.211] has quit [Quit: Leaving.] | 00:57 | |
-!- blackburn [~blackburn@109.226.92.17] has quit [Quit: Leaving.] | 02:55 | |
-!- K0stIa [~kostia@alt2.hk.cvut.cz] has joined #shogun | 06:39 | |
-!- gsomix [~gsomix@95.153.169.120] has joined #shogun | 07:13 | |
-!- n4nd0 [~nando@s83-179-44-135.cust.tele2.se] has joined #shogun | 07:14 | |
-!- K0stIa [~kostia@alt2.hk.cvut.cz] has quit [Quit: Leaving.] | 07:20 | |
-!- K0stIa1 [~kostia@alt2.hk.cvut.cz] has joined #shogun | 07:20 | |
-!- zxtx [~zv@cpe-75-83-151-252.socal.res.rr.com] has joined #shogun | 07:44 | |
-!- K0stIa1 [~kostia@alt2.hk.cvut.cz] has left #shogun [] | 07:53 | |
-!- K0stIa1 [~kostia@alt2.hk.cvut.cz] has joined #shogun | 07:55 | |
-!- K0stIa1 [~kostia@alt2.hk.cvut.cz] has left #shogun [] | 07:55 | |
-!- gsomix [~gsomix@95.153.169.120] has quit [Ping timeout: 264 seconds] | 08:06 | |
-!- uricamic [~uricamic@2001:718:2:1634:6594:2095:459a:894f] has joined #shogun | 08:49 | |
-!- uricamic [~uricamic@2001:718:2:1634:6594:2095:459a:894f] has left #shogun [] | 08:50 | |
-!- uricamic [~uricamic@2001:718:2:1634:6594:2095:459a:894f] has joined #shogun | 08:50 | |
-!- K0stIa [~kostia@2001:718:2:1634:76f0:6dff:fe92:4013] has joined #shogun | 08:52 | |
n4nd0 | sonney2k: around? | 09:26 |
---|---|---|
-!- pluskid [~pluskid@202.130.113.141] has joined #shogun | 09:28 | |
-!- K0stIa [~kostia@2001:718:2:1634:76f0:6dff:fe92:4013] has left #shogun [] | 09:49 | |
-!- K0stIa [~kostia@2001:718:2:1634:76f0:6dff:fe92:4013] has joined #shogun | 09:58 | |
-!- gsomix [~gsomix@95.153.174.248] has joined #shogun | 10:00 | |
K0stIa | Hi, all! I installed shogun for python and Gaussian, GMM are missing in my shogun.Distributions. My linux distribution is ArchLinux, I'm using python2.7. | 10:11 |
K0stIa | can anyone help me with this issue.. | 10:19 |
n4nd0 | K0stIa: hi | 10:21 |
n4nd0 | K0stIa: did you install from packages or from source? | 10:21 |
K0stIa | n4nd0: Hi! from source. | 10:22 |
K0stIa | I took it from github | 10:23 |
pluskid | K0stIa: did you install LaPack? | 10:23 |
n4nd0 | K0stIa: let me check if Gaussian and GMM have some dependencias that you might not be using | 10:23 |
n4nd0 | exactly | 10:24 |
pluskid | K0stIa: AFAIK, lapack in Arch do not have header files, which is needed for shogun to compile some of its components | 10:24 |
pluskid | K0stIa: you will have to either install lapack manually or you can use the package atlas-lapack from AUR | 10:24 |
K0stIa | pluskid: I'm using MKL.... and btw: still shogun didn't find it despite telling it implicitly through --libs, --inludes. | 10:26 |
pluskid | K0stIa: see above, did you install lapack? | 10:27 |
K0stIa | pluskid: no, I didn't. | 10:28 |
K0stIa | pluskid: MKL is not enough ? | 10:28 |
pluskid | K0stIa: I'm sorry, what do you mean by MKL? | 10:28 |
pluskid | did ./configure report that it has detected a valid lapack? | 10:29 |
K0stIa | pluskid: Math Kernel Library | 10:29 |
K0stIa | pluskid: as I said ./configure didn't find it... | 10:29 |
K0stIa | pluskid: however I built numpy with MKL successfully | 10:30 |
pluskid | K0stIa: then I'm not sure | 10:31 |
pluskid | K0stIa: in the ./configure script, there's a section for checking Intel MKL | 10:31 |
pluskid | K0stIa: you might have to modify that to suit your situation :-/ I'm not familiar with this | 10:31 |
pluskid | sonney2k: maybe you can help with this | 10:32 |
n4nd0 | K0stIa: so MKL can be a substitute of lapack for shogun? | 10:33 |
n4nd0 | I didn't know about that ... | 10:33 |
pluskid | n4nd0: there's a section in ./configure that checks many provider of lapack, atlas is one provider (that I'm using), MKL seems to be another | 10:35 |
K0stIa | n4nd0: as far as I know lapack is just some interface and MKL has also realization of it.. at least I built my numpy with it successfully and found it to be much faster then lapack library taken from AUR | 10:35 |
pluskid | but not sure why it fails | 10:35 |
K0stIa | n4nd0: i.e. atlas | 10:35 |
n4nd0 | aham, I had no idea about that | 10:36 |
n4nd0 | pluskid, K0stIa: thanks for the info | 10:37 |
K0stIa | n4nd0: you are wellcome | 10:37 |
pluskid | maybe blackburn can help with this when he shows up | 10:37 |
K0stIa | pluskid: ok, I will wait for him... and I can try on different machine.. I have another lapack there... | 10:38 |
-!- gsomix [~gsomix@95.153.174.248] has quit [Ping timeout: 250 seconds] | 10:46 | |
-!- blackburn [~blackburn@109.226.92.17] has joined #shogun | 11:01 | |
blackburn | K0stIa: hey still struggling with MKL? | 11:03 |
-!- yoo [2eda6d52@gateway/web/freenode/ip.46.218.109.82] has joined #shogun | 11:05 | |
K0stIa | blackburn: sort of... | 11:05 |
yoo | hi all | 11:05 |
blackburn | K0stIa: did you add --includes and --libs flags? | 11:06 |
K0stIa | blackburn: yes I did | 11:06 |
blackburn | still not detected? | 11:06 |
K0stIa | blackburn: yes | 11:06 |
blackburn | okay when what did you add in there? | 11:07 |
K0stIa | blackburn: something like ./configure --interfaces=python_modular \ | 11:09 |
K0stIa | --prefix=/usr \ | 11:09 |
K0stIa | --libs=/opt/intel/mkl/lib/intel64 \ | 11:09 |
K0stIa | --includes=/opt/intel/mkl/include | 11:09 |
koen-shogun | you should check configure.log , on what happened in the MKL checks | 11:09 |
blackburn | koen-shogun: did you ever link shogun on mkl? | 11:10 |
koen-shogun | it probably fails to compile (or to link) there | 11:10 |
koen-shogun | no, but I had to fix the same issues for ATLAS on my machine | 11:10 |
K0stIa | MKL is missing in my log... | 11:10 |
koen-shogun | go to the "============ Checking for Intel MKL support ============" in configure.log | 11:10 |
koen-shogun | and copy the text underneath that :) | 11:11 |
koen-shogun | until the part where it says "Result is: no | 11:11 |
koen-shogun | ##########################################" | 11:11 |
K0stIa | koen-shogun: ok | 11:11 |
koen-shogun | oh, and which version of MKL do you have? | 11:12 |
K0stIa | ============ Checking for Intel MKL support ============ | 11:14 |
K0stIa | #include <mkl_cblas.h> | 11:14 |
K0stIa | void dpotrf(char* uplo, int* n, double* a, int* lda, int* info); | 11:14 |
K0stIa | int main(int argc, char** argv) | 11:14 |
K0stIa | { | 11:14 |
-!- K0stIa [~kostia@2001:718:2:1634:76f0:6dff:fe92:4013] has quit [Excess Flood] | 11:14 | |
-!- K0stIa [~kostia@2001:718:2:1634:76f0:6dff:fe92:4013] has joined #shogun | 11:14 | |
pluskid | K0stIa: use some external service to paste large chunk of code :) | 11:15 |
K0stIa | pluskid: for instance ? :) | 11:15 |
pluskid | like gist from github | 11:15 |
K0stIa | ok | 11:15 |
K0stIa | :) | 11:15 |
K0stIa | koen-shogun: pluskid: https://gist.github.com/3205809 | 11:16 |
K0stIa | koen-shogun: about version of mkl it's 2011_sp1.9.293 | 11:18 |
koen-shogun | it finds your include files ok | 11:19 |
koen-shogun | but it fails on linking | 11:19 |
K0stIa | koen-shogun: you mean I have to change --libs parameter? | 11:26 |
-!- n4nd0 [~nando@s83-179-44-135.cust.tele2.se] has quit [Quit: leaving] | 11:28 | |
blackburn | K0stIa: I actually expect something is wrong with -l* | 11:28 |
K0stIa | blackburn: then it's not my problem, right ? | 11:30 |
koen-shogun | well, but you can fix it | 11:30 |
koen-shogun | the Intel MKL changes the libraries to link against every other version | 11:30 |
blackburn | K0stIa: can you paste somewhere contents of /opt/intel/mkl/lib/intel64 please? | 11:30 |
K0stIa | blackburn: ok | 11:30 |
koen-shogun | for example, I have a link line (for some other program) that looks like -lmkl_intel_ilp64 -lmkl_intel_thread -lmkl_core | 11:31 |
koen-shogun | oh, and the complete needed link line then needs "-Wl,--start-group -lmkl_intel_ilp64 -lmkl_intel_thread -lmkl_core -Wl,--end-group" | 11:32 |
K0stIa | blackburn: https://gist.github.com/3205860 | 11:32 |
blackburn | koen-shogun: I expect thing we have for mkl now is outdated | 11:32 |
koen-shogun | if you edit configure, go to "echocheck "Intel MKL support"" | 11:33 |
koen-shogun | and replace -lmkl -lguide -lmkl_lapack twice with for example my link line, that might work | 11:33 |
K0stIa | koen-shogun: thanks, I will try... | 11:34 |
blackburn | K0stIa: lines 3168,3174,3176,3182 | 11:34 |
blackburn | try to replace it with | 11:34 |
blackburn | -Wl,--start-group -lmkl_intel_ilp64 -lmkl_intel_thread -lmkl_core -Wl,--end-group | 11:34 |
blackburn | and run configure | 11:34 |
K0stIa | blackburn: still the same :( | 11:37 |
koen-shogun | what does configure.log say now? | 11:39 |
K0stIa | koen-shogun: https://gist.github.com/3205900 | 11:42 |
blackburn | interesting | 11:42 |
blackburn | that's openmp missing I guess | 11:42 |
blackburn | K0stIa: either replace intel_thread with sequential | 11:43 |
blackburn | or add -fopenmp | 11:43 |
koen-shogun | do you also use the intel compiler? | 11:44 |
koen-shogun | because then it's -openmp | 11:44 |
K0stIa | koen-shogun: no, gcc | 11:44 |
koen-shogun | ok, and do you have a iomp5 library somewhere? then you'd need -liomp5 (plus an additional --libs path) | 11:45 |
koen-shogun | maybe sequential is the easiest solution | 11:45 |
K0stIa | ok, I have to go now... I will come back in 30-40 minutes, then I will inform you about my situation. koen-shogun, blackburn thanks for helping | 11:47 |
blackburn | K0stIa: ok I will be around | 11:48 |
CIA-18 | shogun: Sergey Lisitsyn master * r2fd15a1 / src/configure : Updated MKL linking flags - https://github.com/shogun-toolbox/shogun/commit/2fd15a161921a57c0dccd6db2bf1188e0d2829cc | 12:05 |
-!- needsch [~user@ip-176-198-229-75.unitymediagroup.de] has joined #shogun | 12:10 | |
-!- emrecelikten [~emre@176.40.234.211] has joined #shogun | 12:32 | |
K0stIa | blackburn: I see someone updated ./configure, so I updated shogun from github, and tried to install it, it failed on make command. this is what it says: http://pastebin.com/6U8t2brg | 13:07 |
blackburn | K0stIa: well at least it detected mkl :) | 13:08 |
K0stIa | yes, that's true :) | 13:08 |
blackburn | K0stIa: can you paste /opt/intel/mkl/include/mkl_cblas.h somewhere? | 13:09 |
K0stIa | blackburn: yes | 13:10 |
K0stIa | blackburn: http://pastebin.com/Qi6SF10x | 13:12 |
blackburn | ohh | 13:13 |
blackburn | K0stIa: please update and try again | 13:14 |
CIA-18 | shogun: Sergey Lisitsyn master * rda74d42 / src/shogun/mathematics/lapack.h : Removed wrong enum keywords in lapack wrappers - http://git.io/gXZgGA | 13:14 |
K0stIa | blackburn: ok | 13:15 |
-!- n4nd0 [~nando@s83-179-44-135.cust.tele2.se] has joined #shogun | 13:23 | |
blackburn | K0stIa: so did that help | 13:29 |
K0stIa | blackburn: http://pastebin.com/BTfxRtBv | 13:30 |
blackburn | okay got it | 13:30 |
CIA-18 | shogun: Sergey Lisitsyn master * r77433fe / src/shogun/mathematics/lapack.cpp : Removed one more wrong enum in lapack - http://git.io/kJFZug | 13:31 |
blackburn | K0stIa: hopefully should work now | 13:31 |
K0stIa | blackburn: ok, I'll check | 13:31 |
-!- pluskid [~pluskid@202.130.113.141] has quit [Ping timeout: 272 seconds] | 13:52 | |
-!- pluskid [~pluskid@111.120.36.210] has joined #shogun | 14:14 | |
-!- yoo [2eda6d52@gateway/web/freenode/ip.46.218.109.82] has quit [Quit: Page closed] | 14:21 | |
n4nd0 | yaaay!!! \o/ | 14:27 |
-!- yoo [2eda6d52@gateway/web/freenode/ip.46.218.109.82] has joined #shogun | 14:28 | |
n4nd0 | I think I can say with confidence that my hm-svm implementation is correct! | 14:28 |
yoo | n4nd0: where can we find the implementation of hm-svm with shogun ? | 14:34 |
n4nd0 | yoo: it is in my branch right now | 14:35 |
n4nd0 | yoo: I have to polished some things - lot of debugging code currently | 14:35 |
n4nd0 | yoo: do you want to use it? | 14:35 |
yoo | n4nd0: it seems interesting :) | 14:35 |
n4nd0 | yoo: I will probably prepare a pull request today | 14:36 |
yoo | n4nd0: I wanted to code one myself, then seeing yours would be cool. I am not enough still good with shogun anyway .. | 14:36 |
yoo | n4nd0: nice: | 14:37 |
n4nd0 | yoo: so have you experience with hm-svm? | 14:39 |
yoo | yea but code is creepy .. looks like hybrid svm-hmm , I thought to rewrite something in the near future | 14:40 |
yoo | but since I have discovered your work in shogun, I would like to better understand the codes and mb contribute then. | 14:42 |
yoo | I still have some bp understand some partsof the code architecture | 14:43 |
n4nd0 | yoo: do you know about applications of it? | 14:46 |
n4nd0 | yoo: I am looking for something appealing to present in my thesis | 14:46 |
yoo | I will take a look if you want .. no explosive idea in mind right now. | 14:50 |
yoo | let me know if PR today ! + | 14:51 |
yoo | what is the subject of your thesis btw ? | 14:52 |
n4nd0 | yoo: completely open right now | 14:57 |
n4nd0 | nothing decided | 14:57 |
n4nd0 | I just like the idea of using the work I have been doing during this summer since I feel that I've learnt interesting things | 14:58 |
n4nd0 | + the ones I am still to learn :) | 14:58 |
emrecelikten | n4nd0: Congratulations | 15:02 |
emrecelikten | :) | 15:02 |
n4nd0 | emrecelikten: thanks! | 15:06 |
CIA-18 | shogun: Sergey Lisitsyn master * r1cee68b / (13 files in 3 dirs): Multitask crossvalidation support - http://git.io/hjS24Q | 15:11 |
-!- K0stIa [~kostia@2001:718:2:1634:76f0:6dff:fe92:4013] has left #shogun [] | 15:29 | |
-!- pluskid [~pluskid@111.120.36.210] has quit [Ping timeout: 248 seconds] | 15:58 | |
-!- pluskid [~pluskid@202.130.113.141] has joined #shogun | 15:59 | |
-!- gsomix [~gsomix@95.153.179.239] has joined #shogun | 16:01 | |
-!- pluskid [~pluskid@202.130.113.141] has quit [Quit: Leaving] | 16:37 | |
-!- heiko1 [~heiko@host86-181-156-213.range86-181.btcentralplus.com] has joined #shogun | 16:38 | |
* sonney2k ohh man it feels so good - true broadband internet :D | 16:40 | |
@sonney2k | gsomix, how are things? | 16:41 |
heiko1 | sonney2k, so you are finally back? :) | 16:41 |
@sonney2k | heiko1, yes I am | 16:43 |
heiko1 | hope you had a good time | 16:43 |
-!- ckwidmer [8ca3fe9d@gateway/web/freenode/ip.140.163.254.157] has joined #shogun | 16:44 | |
gsomix | sonney2k, still working on buffers for SG. | 16:44 |
gsomix | I have some problems with it. Pointers crashes. | 16:45 |
@sonney2k | gsomix, lets talk tonight or tomorrow | 16:45 |
gsomix | ok, tonight | 16:46 |
gsomix | I can swim in the sea before. :D | 16:46 |
@sonney2k | heiko1, yes indeed :) | 16:46 |
gsomix | sonney2k, did you receive my letters? | 16:47 |
-!- n4nd0 [~nando@s83-179-44-135.cust.tele2.se] has quit [Quit: leaving] | 16:49 | |
-!- gsomix [~gsomix@95.153.179.239] has quit [Ping timeout: 244 seconds] | 17:24 | |
blackburn | sonney2k: yes, did you receive his love letters | 17:46 |
blackburn | ? | 17:46 |
blackburn | :D | 17:46 |
CIA-18 | shogun: Heiko Strathmann master * re532b28 / (5 files in 2 dirs): Merge pull request #680 from karlnapf/master (+6 more commits...) - http://git.io/4sj-9g | 17:46 |
-!- uricamic [~uricamic@2001:718:2:1634:6594:2095:459a:894f] has quit [Quit: Leaving.] | 17:56 | |
-!- tiger_eye [~tiger_eye@rrcs-24-106-116-254.central.biz.rr.com] has left #shogun ["Leaving"] | 18:05 | |
-!- puffin444 [62e3926e@gateway/web/freenode/ip.98.227.146.110] has joined #shogun | 18:07 | |
CIA-18 | shogun: Heiko Strathmann master * rda0380e / src/shogun/statistics/LinearTimeMMD.cpp : added threshold computation using inverse gaussian cdf - http://git.io/iBOAwg | 18:35 |
CIA-18 | shogun: Heiko Strathmann master * rb762388 / (4 files in 3 dirs): Merge pull request #681 from karlnapf/master - http://git.io/dndEcw | 18:35 |
CIA-18 | shogun: Heiko Strathmann master * rf3492a1 / (3 files in 2 dirs): added inverse gaussian cdf plus tests - http://git.io/gRkyfw | 18:35 |
-!- yoo [2eda6d52@gateway/web/freenode/ip.46.218.109.82] has quit [Quit: Page closed] | 18:37 | |
-!- blackburn [~blackburn@109.226.92.17] has quit [Read error: Connection reset by peer] | 19:10 | |
-!- puffin444_ [62e3926e@gateway/web/freenode/ip.98.227.146.110] has joined #shogun | 19:12 | |
-!- puffin444 [62e3926e@gateway/web/freenode/ip.98.227.146.110] has quit [Ping timeout: 245 seconds] | 19:13 | |
-!- blackburn [~blackburn@109.226.100.87] has joined #shogun | 19:24 | |
-!- heiko1 [~heiko@host86-181-156-213.range86-181.btcentralplus.com] has left #shogun [] | 19:31 | |
-!- n4nd0 [~nando@s83-179-44-135.cust.tele2.se] has joined #shogun | 20:28 | |
n4nd0 | hi there | 20:28 |
n4nd0 | sonney2k: around? | 20:29 |
n4nd0 | blackburn: do you know if sonney2k is/will be around here this evening? | 20:29 |
blackburn | n4nd0: I only know that he was online tomorrow :) | 20:30 |
n4nd0 | blackburn: ehh ok | 20:30 |
n4nd0 | like he will be tomorrow right? | 20:30 |
blackburn | ehm | 20:32 |
blackburn | n4nd0: that's crazy but this happens with me for third time | 20:32 |
blackburn | tomorrow = today | 20:32 |
blackburn | :D | 20:32 |
n4nd0 | haha ok :) | 20:33 |
blackburn | no idea what makes me mix these words | 20:33 |
n4nd0 | don't worry, these things happen | 20:33 |
n4nd0 | something I talk to myself | 20:33 |
n4nd0 | n4nd0: like this | 20:33 |
n4nd0 | :D | 20:33 |
-!- n4nd0 [~nando@s83-179-44-135.cust.tele2.se] has quit [Ping timeout: 246 seconds] | 20:53 | |
-!- gsomix [~gsomix@95.153.161.248] has joined #shogun | 21:39 | |
-!- gsomix [~gsomix@95.153.161.248] has quit [Quit: Ex-Chat] | 21:50 | |
-!- gsomix [~gsomix@95.153.161.248] has joined #shogun | 21:51 | |
-!- n4nd0 [~nando@s83-179-44-135.cust.tele2.se] has joined #shogun | 21:52 | |
gsomix | good evening | 21:52 |
-!- n4nd0 [~nando@s83-179-44-135.cust.tele2.se] has quit [Ping timeout: 255 seconds] | 21:59 | |
@sonney2k | gsomix, I am around just now... | 22:02 |
@sonney2k | gsomix, so what is the trouble with SG* ? | 22:02 |
@sonney2k | that these are not pointers? | 22:03 |
@sonney2k | gsomix, and yes I received your emails so I didn't have to worry :) | 22:04 |
@sonney2k | gsomix, I hope that you can fix this up within the next few days such that we can work on the model selection typemaps... | 22:08 |
-!- n4nd0 [~nando@s83-179-44-135.cust.tele2.se] has joined #shogun | 22:11 | |
blackburn | n4nd0: sonney2k | 22:11 |
blackburn | have a good date | 22:12 |
blackburn | :D | 22:12 |
@sonney2k | blackburn, what? | 22:12 |
@sonney2k | n4nd0, ??? | 22:12 |
blackburn | sonney2k: n4nd0 missed you | 22:13 |
@sonney2k | heh | 22:13 |
@sonney2k | I will go to bed soon but just now I am around :) | 22:13 |
n4nd0 | sonney2k: yes, I was writing my weekly report right now making you that question | 22:13 |
blackburn | like the desert misses the rain | 22:13 |
n4nd0 | I will ask now instead | 22:13 |
@sonney2k | blackburn, :P | 22:13 |
n4nd0 | sonney2k: so now according to the timetable I should work on CRFs | 22:13 |
n4nd0 | sonney2k: but Georg cannot help me with that | 22:14 |
n4nd0 | Nico is still on holidays | 22:14 |
@sonney2k | n4nd0, does Georg have some other ideas? | 22:14 |
n4nd0 | and we didn't talk about how they should fit into our SO framework | 22:14 |
blackburn | n4nd0: may be you just hang around with guys, have some beer? | 22:14 |
@sonney2k | I would expect CRFs to not be that different | 22:14 |
n4nd0 | sonney2k: yes, there is some things in the hmsvm toolbox that are not currently in my implementation | 22:15 |
n4nd0 | like the plifs | 22:15 |
@sonney2k | n4nd0, and then when nico comes back continue to do CRFs ? | 22:15 |
@sonney2k | n4nd0, plifs are some important thing - and they are in shogun already | 22:15 |
n4nd0 | sonney2k: I am afraid there won't be enough time to finish it before gsoc finishes | 22:15 |
blackburn | mind to continue? ;) | 22:15 |
@sonney2k | so certainly a nice replacement to do instead of CRFs for now | 22:15 |
n4nd0 | sonney2k: I know how plifs work in the hmsvm toolbox, and I think it won't take much time to add them | 22:16 |
@sonney2k | n4nd0, well we all hope you stick around after gsoc ... | 22:16 |
n4nd0 | blackburn, sonney2k: sure I can work on that later | 22:16 |
n4nd0 | no problem for me | 22:16 |
blackburn | n4nd0: life is ppain | 22:16 |
n4nd0 | but for Nico? | 22:16 |
blackburn | gsoc is hell | 22:16 |
@sonney2k | n4nd0, look at the CPlif classes | 22:16 |
@sonney2k | they should do the plif business | 22:16 |
blackburn | you are all tied to shogun support for ever | 22:16 |
n4nd0 | sonney2k: ok, I will prepare the pull request and later do that | 22:16 |
n4nd0 | blackburn: :) | 22:17 |
@sonney2k | blackburn, one shogun to rule them all or so :D | 22:17 |
@sonney2k | n4nd0, I would even say that adding plifs is more important than CRFs | 22:17 |
n4nd0 | sonney2k: ok | 22:17 |
blackburn | how to use plifs? | 22:17 |
n4nd0 | sonney2k: but I mean, if I do it as they are done in the hmsvm toolbox | 22:17 |
blackburn | it took 1.5 years for me to get what does it mean | 22:18 |
n4nd0 | sonney2k: it is not much code | 22:18 |
@sonney2k | so if you can do the hmsvm toolbox implementation in shogun too - getting same result as hmsvm toolbox that would be even better | 22:18 |
n4nd0 | sonney2k: right now I get the exact same results for integer data :) | 22:18 |
@sonney2k | blackburn, for what? | 22:18 |
n4nd0 | I am very happy with that | 22:18 |
@sonney2k | n4nd0, thats a start :D | 22:18 |
blackburn | sonney2k: how to use plifs for anything? | 22:18 |
* sonney2k too | 22:18 | |
@sonney2k | blackburn, ahh well consider that a linear function is not enough | 22:19 |
n4nd0 | sonney2k: believe me, it's more than that. At least in the hmsvm toolbox the code that is for the plifs changes very little compare to the other one | 22:19 |
@sonney2k | like svm | 22:19 |
@sonney2k | so you use a piece wise linear function | 22:19 |
n4nd0 | with the other one I mean just working with discrete observations | 22:19 |
blackburn | sonney2k: but is plif a model or an output? | 22:19 |
n4nd0 | blackburn: it is part of the model | 22:19 |
@sonney2k | blackburn, not a model - just some transformation | 22:20 |
n4nd0 | oops | 22:20 |
@sonney2k | well or part of the model :D | 22:20 |
n4nd0 | :) | 22:20 |
blackburn | so it just to support some broken hyperplanes? | 22:20 |
@sonney2k | for svms you can emulate that by using these binneddotfeatures | 22:20 |
n4nd0 | sonney2k: and what about the coffin idea in the SO framework? | 22:21 |
blackburn | sonney2k: I never thought of that mean of binned dot | 22:21 |
@sonney2k | n4nd0, I thought you are doing it already - no? | 22:21 |
n4nd0 | sonney2k: mmm no | 22:21 |
@sonney2k | blackburn, binned dot is making the svm learn a piecewise linearfunction | 22:22 |
@sonney2k | with fixed 'stuetzstellen; | 22:22 |
n4nd0 | Nico told me I should talk to you about that once I had nothing else to do :D | 22:22 |
blackburn | sonney2k: I got it already | 22:22 |
@sonney2k | n4nd0, hmmhh | 22:22 |
blackburn | no f idea what is stuetzstellen | 22:22 |
blackburn | sounds like luftwaffe | 22:22 |
blackburn | or panzerkrieg | 22:22 |
@sonney2k | blackburn, supporting points or so | 22:22 |
@sonney2k | n4nd0, I was actually hoping that you can do that too... | 22:23 |
@sonney2k | ok gtg for now | 22:23 |
@sonney2k | will hopefully be around tomorrow too. | 22:23 |
n4nd0 | sonney2k: let's talk about it tomorrow then? | 22:23 |
blackburn | n4nd0: can I help you with coffinization? | 22:23 |
blackburn | I've got some experience with it | 22:23 |
n4nd0 | blackburn: nice, that would be great | 22:24 |
blackburn | put your code into the coffin ya know | 22:24 |
n4nd0 | blackburn, sonney2k: I will focus on that after the PLiFs, ok? | 22:24 |
blackburn | focus on what? | 22:24 |
blackburn | n4nd0: guide me to some solver code | 22:25 |
n4nd0 | on the coffinization | 22:25 |
n4nd0 | blackburn: what do you mean? | 22:25 |
n4nd0 | you want to see the code of my solver I understand :) | 22:25 |
blackburn | n4nd0: well I mean I could take a look to the code | 22:25 |
n4nd0 | sure | 22:25 |
n4nd0 | blackburn: https://github.com/iglesias/shogun/blob/so/src/shogun/structure/PrimalMosekSOSVM.cpp | 22:26 |
n4nd0 | I am sorry for the debug traces ... they are going out soon | 22:26 |
n4nd0 | train_machine | 22:26 |
blackburn | n4nd0: I will never forgive you | 22:27 |
blackburn | :D | 22:27 |
n4nd0 | I guess that the parts you are interested in are the ones related to the use of features | 22:27 |
gsomix | sonney2k, hey | 22:27 |
n4nd0 | argmax function and compute_loss_arg | 22:27 |
gsomix | oh, ok :) | 22:28 |
blackburn | n4nd0: well compute_loss_arg is not really | 22:30 |
blackburn | coffinizable | 22:30 |
n4nd0 | blackburn: tell me why | 22:30 |
n4nd0 | we should be using an special kind of features I am afraid | 22:30 |
blackburn | n4nd0: I can't see where do you use features there.. | 22:30 |
n4nd0 | something that has access both to the labels and features | 22:30 |
blackburn | are we talking about the same? | 22:30 |
n4nd0 | the problem is that in SO learning | 22:30 |
n4nd0 | the vecotr that goes in the product with w looks like this | 22:31 |
n4nd0 | Psi(x,y) | 22:31 |
n4nd0 | x is a feature vector | 22:31 |
n4nd0 | y is a label | 22:31 |
blackburn | so where do we have explicit access to data? | 22:32 |
n4nd0 | so to make it coffinizable I think it is required to use a special type of features for that | 22:32 |
n4nd0 | in the models | 22:32 |
n4nd0 | https://github.com/iglesias/shogun/blob/so/src/shogun/structure/HMSVMModel.cpp | 22:33 |
n4nd0 | CHMSVMModel::get_joint_feature_vector | 22:33 |
blackburn | checking | 22:34 |
blackburn | pretty complex thing huh | 22:36 |
blackburn | n4nd0: from what I see now it is not really needed to make it use add and dot | 22:36 |
n4nd0 | what arguments for these add and dot? | 22:37 |
blackburn | n4nd0: basically most solvers work using two operations | 22:38 |
blackburn | add some train feature vector to some given vector | 22:38 |
blackburn | and compute dot product of some train feature vector and given vector | 22:39 |
n4nd0 | ok | 22:39 |
n4nd0 | that is done in compute_loss_arg | 22:39 |
n4nd0 | here the train feature vector are psi_pred and psi_truth | 22:40 |
n4nd0 | just the dot product actually | 22:40 |
blackburn | yeah | 22:40 |
blackburn | that is not the point of coffin | 22:40 |
blackburn | main point is that you don't need to have vectors itself explicitly | 22:41 |
n4nd0 | yes, I understood that when I read the paper | 22:41 |
n4nd0 | I didn't get though why it is so benefecious | 22:42 |
n4nd0 | and how to do it for SO | 22:42 |
blackburn | n4nd0: first is easy question | 22:42 |
blackburn | second is not :) | 22:42 |
blackburn | n4nd0: okay imagine you have a set of images | 22:42 |
blackburn | say 1M of images | 22:42 |
blackburn | it is a common way to extend dataset using some transformations | 22:43 |
blackburn | rotations may be | 22:43 |
blackburn | so if you don't need to have all images explicitly | 22:44 |
blackburn | and if you have some procedure to compute dot on transformed images as efficiently | 22:44 |
blackburn | you easily extend your train set but your memory requirements do not raise up | 22:45 |
n4nd0 | ok | 22:45 |
blackburn | or other example - you have fast preprocessing routine (homogeneous kernel map is very fast for example) | 22:45 |
blackburn | by default it makes d = 3*d | 22:45 |
blackburn | so you need to have 3x more memory | 22:45 |
blackburn | n4nd0: see what I mean? | 22:46 |
n4nd0 | but it is something that takes more time than having everything stored in memory right? I understand that it is beneficious in any case because for large datasets everything won't fit in memory simply | 22:46 |
n4nd0 | blackburn: yes, but ^ | 22:46 |
n4nd0 | just to be sure | 22:46 |
blackburn | n4nd0: yes that can require more time | 22:47 |
n4nd0 | for the rotations example | 22:47 |
blackburn | not the best example probably | 22:47 |
n4nd0 | the same rotation will be done more than once for example | 22:47 |
n4nd0 | :D | 22:47 |
blackburn | what if we just cut | 22:47 |
blackburn | various sliding window positions | 22:47 |
blackburn | you need to have only one image and it can represent 100 vectors | 22:48 |
blackburn | with the same memory req | 22:48 |
blackburn | you just compute dot with different initial position or so | 22:48 |
n4nd0 | I didn't get this example | 22:49 |
blackburn | 0 0 0 | 22:49 |
blackburn | 0 0 0 | 22:49 |
blackburn | 0 0 0 | 22:49 |
blackburn | ^ image | 22:49 |
blackburn | 1 1 0 | 22:49 |
blackburn | 1 1 0 | 22:49 |
blackburn | 0 0 0 | 22:49 |
blackburn | first window position | 22:49 |
blackburn | 0 1 1 | 22:49 |
blackburn | 0 1 1 | 22:49 |
blackburn | 0 0 0 | 22:49 |
blackburn | second.. | 22:50 |
blackburn | so on | 22:50 |
n4nd0 | ok | 22:50 |
blackburn | n4nd0: it is not that costly to compute dot or add | 22:50 |
blackburn | with just a region of image available | 22:50 |
blackburn | n4nd0: sparse uses it too | 22:53 |
n4nd0 | I don't think I understand the real magic of coffin | 22:53 |
-!- gsomix [~gsomix@95.153.161.248] has quit [Ping timeout: 264 seconds] | 22:53 | |
blackburn | n4nd0: hmmm I am in search of good example :D | 22:56 |
n4nd0 | :) | 22:57 |
n4nd0 | maybe I am just looking for an easter egg | 22:57 |
n4nd0 | sonney2k told me that COFFIN was quite relevant | 22:57 |
blackburn | n4nd0: well it is actually a good soft engineering example as well | 22:58 |
blackburn | you don't have to dispatch what type of features you are working on | 22:58 |
blackburn | you just make features do that for you | 22:58 |
blackburn | so you don't care about storage/preprocessing/anytihng | 22:58 |
blackburn | if you work with sparse vectors | 22:59 |
blackburn | you can't get pointer to dense vector to add with | 22:59 |
blackburn | you have to handle it somehow everywhere | 22:59 |
blackburn | with coffin you don't care | 22:59 |
CIA-18 | shogun: Sergey Lisitsyn master * r7893db8 / (3 files): Removed multitask subset support - http://git.io/YOrJPA | 23:03 |
-!- ckwidmer [8ca3fe9d@gateway/web/freenode/ip.140.163.254.157] has quit [Ping timeout: 245 seconds] | 23:09 | |
-!- zxtx [~zv@cpe-75-83-151-252.socal.res.rr.com] has quit [Ping timeout: 244 seconds] | 23:10 | |
-!- zxtx [~zv@cpe-75-83-151-252.socal.res.rr.com] has joined #shogun | 23:12 | |
--- Log closed Tue Jul 31 00:00:17 2012 |
Generated by irclog2html.py 2.10.0 by Marius Gedminas - find it at mg.pov.lt!