--- Log opened Thu Oct 04 00:00:17 2012 | ||
blackburn | okay getting late adoniscik, heiko see you later | 00:04 |
---|---|---|
adoniscik | didn't work sorry! | 00:04 |
heiko | take care blackburn | 00:04 |
adoniscik | one sec if you will | 00:04 |
blackburn | adoniscik: what's up? | 00:05 |
adoniscik | /usr/local/lib/libnlopt.a: could not read symbols: Bad value | 00:05 |
adoniscik | does that mean the make install did not work properly? | 00:05 |
blackburn | is that after recompiling? | 00:05 |
adoniscik | is it ok to compile in $home and move the whole folder afterwards? | 00:05 |
adoniscik | I compiled with sudo in /usr/lib | 00:06 |
blackburn | adoniscik: did you make "sudo make install" afterwards? | 00:06 |
adoniscik | yes, I did | 00:06 |
blackburn | no need to put it anywhere, make install should take care of that | 00:06 |
adoniscik | I forgot this message: | 00:07 |
adoniscik | /usr/bin/ld: /usr/local/lib/libnlopt.a(optimize.o): relocation R_X86_64_32S against `.text' can not be used when making a shared object; recompile with -fPIC | 00:07 |
blackburn | is it still there? | 00:07 |
adoniscik | just to be clear, I recompiled nlopt in its own folder | 00:08 |
adoniscik | is that what you meant? | 00:08 |
blackburn | what do you mean its own folder? | 00:08 |
adoniscik | ~/nlopt-2.3 | 00:08 |
blackburn | yeah, | 00:08 |
blackburn | then you "make" it | 00:08 |
blackburn | and then "sudo make install" | 00:08 |
blackburn | right? | 00:08 |
adoniscik | yes | 00:08 |
blackburn | did you run configure once again afterwards? | 00:09 |
adoniscik | it says it installed in /usr/local/lib | 00:09 |
adoniscik | configure for shogun? with enabled-shared? | 00:09 |
blackburn | yes, but without enabled-shared | 00:10 |
blackburn | however | 00:10 |
blackburn | just try make clean | 00:10 |
blackburn | in shogun | 00:10 |
blackburn | and make it again | 00:10 |
adoniscik | after configure I run make clean? | 00:10 |
blackburn | yeah that would work | 00:10 |
blackburn | or even without configure | 00:10 |
adoniscik | done | 00:10 |
adoniscik | what next? | 00:10 |
blackburn | make | 00:10 |
blackburn | btw I suggest to use make -j2 or make -j4 | 00:11 |
blackburn | to employ more threads | 00:11 |
adoniscik | oh nice. shouldn't that be autodetected too? | 00:11 |
adoniscik | I think my 3770 supports 8 threads | 00:11 |
blackburn | I don't know - it is autotools thing | 00:12 |
blackburn | then make -j8 | 00:12 |
adoniscik | thanks for your patience | 00:12 |
blackburn | 3.4 GHz? that's a big boy | 00:12 |
adoniscik | I'll be happy to get this thing running | 00:12 |
adoniscik | When you need to crunch numbers, do it right... | 00:12 |
blackburn | I've got sandy bridge i5 here | 00:13 |
blackburn | but I have a laptop :D | 00:13 |
adoniscik | do you have experience with GPGPU in shogun? | 00:13 |
blackburn | well we had some experiments before this year gsoc | 00:13 |
adoniscik | I was considering getting a 660Ti with 1500 cores | 00:13 |
adoniscik | in due time | 00:13 |
blackburn | one student was suggesting to GPU us a little | 00:13 |
blackburn | the problem is that for example I have no machine with GPU at all | 00:14 |
adoniscik | so you can't develop shogun for it | 00:14 |
blackburn | neither other developers do | 00:14 |
adoniscik | so you need a donation... | 00:15 |
adoniscik | why not post kickstarter and let people pitch in a few bucks? | 00:15 |
blackburn | heh not really the money thing for me | 00:15 |
blackburn | I just don't need it yet or so | 00:15 |
blackburn | we are all developing shogun for our needs usually | 00:16 |
adoniscik | I assumed it would greatly speed up things, but I don't know so that's why I'm asking. | 00:16 |
blackburn | yeah that would speed up most things | 00:16 |
blackburn | but requires quite a lot of effort still | 00:16 |
blackburn | and some of codes can't be speeded up with GPUs | 00:17 |
blackburn | like string things | 00:17 |
adoniscik | after a long compile I got the same error: /usr/bin/ld: /usr/local/lib/libnlopt.a(optimize.o): relocation R_X86_64_32S against `.text' can not be used when making a shared object; recompile with -fPIC | 00:17 |
adoniscik | /usr/local/lib/libnlopt.a: could not read symbols: Bad value | 00:17 |
blackburn | oh that's bad | 00:17 |
adoniscik | I must be doing something wrong with compiling nlopt | 00:17 |
adoniscik | I assume | 00:17 |
adoniscik | is there a way to skip all the unnecessary steps in compiling shogun and just redo this part? | 00:17 |
blackburn | which part? | 00:18 |
adoniscik | the last one! It's a ridiculously long line: http://pastebin.com/5Tjpxf9z | 00:18 |
blackburn | heh | 00:19 |
shogun-buildbot | build #443 of deb2 - static_interfaces is complete: Success [build successful] Build details are at http://www.shogun-toolbox.org/buildbot/builders/deb2%20-%20static_interfaces/builds/443 | 00:19 |
blackburn | it picks up all the .o filess | 00:19 |
blackburn | into a libray | 00:19 |
blackburn | adoniscik: where did you put nlopt first time? | 00:20 |
adoniscik | in /usr/lib. now it's in $home | 00:21 |
adoniscik | and claims to have installed in /usr/local/lib | 00:21 |
adoniscik | Libraries have been installed in: | 00:21 |
adoniscik | /usr/local/lib | 00:21 |
adoniscik | If you ever happen to want to link against installed libraries | 00:21 |
adoniscik | in a given directory, LIBDIR, you must either use libtool, and | 00:21 |
adoniscik | specify the full pathname of the library, or use the `-LLIBDIR' | 00:21 |
adoniscik | flag during linking and do at least one of the following: | 00:21 |
adoniscik | - add LIBDIR to the `LD_LIBRARY_PATH' environment variable | 00:21 |
adoniscik | during execution | 00:21 |
adoniscik | - add LIBDIR to the `LD_RUN_PATH' environment variable | 00:21 |
adoniscik | during linking | 00:21 |
adoniscik | - use the `-Wl,-rpath -Wl,LIBDIR' linker flag | 00:21 |
adoniscik | - have your system administrator add LIBDIR to `/etc/ld.so.conf' | 00:21 |
adoniscik | that's what make said | 00:21 |
blackburn | so you still have that in /usr/lib? | 00:21 |
adoniscik | no, $home | 00:22 |
blackburn | did you remove it from the /usr/lib? | 00:22 |
adoniscik | yes, I moved the nlopt folder to home | 00:22 |
adoniscik | in case of permission issues during compilation | 00:23 |
adoniscik | not sure where it really belongs, just tell me | 00:23 |
blackburn | okay, thinking | 00:23 |
blackburn | strange it fails only on linking then | 00:24 |
blackburn | adoniscik: okay lets run configure again | 00:25 |
blackburn | of shogun | 00:25 |
adoniscik | any new flags? | 00:25 |
adoniscik | we haven't changed anything though | 00:25 |
adoniscik | or I haven't :) | 00:25 |
blackburn | no, just ./configure --interfaces=python_modular | 00:25 |
blackburn | could you please check it detects nlopt now | 00:26 |
adoniscik | yes for NLOPT support | 00:26 |
adoniscik | I did get it to properly configure python_modular before | 00:26 |
adoniscik | the problem was making it afterwards | 00:26 |
blackburn | yeah I understand | 00:27 |
blackburn | lets switch to nlopt then | 00:27 |
adoniscik | this is the conf result: http://pastebin.com/yvW0QU4k | 00:28 |
adoniscik | with enable shared | 00:28 |
adoniscik | this is sudo make install result http://pastebin.com/1NxGEzWJ | 00:28 |
blackburn | do you have /usr/lib/libnlopt.a file? | 00:30 |
blackburn | without loacl | 00:30 |
blackburn | local* | 00:30 |
adoniscik | no | 00:30 |
adoniscik | bingo? | 00:30 |
blackburn | no, would be if there was | 00:30 |
blackburn | I am trying to find a reason | 00:31 |
blackburn | why it uses old nlopt | 00:31 |
adoniscik | I have two versions? | 00:31 |
blackburn | yeah I thought so but it seems that no | 00:31 |
adoniscik | locate finds only one; in /usr/local/lib | 00:32 |
blackburn | trying to compile with nlopt enabled now | 00:33 |
adoniscik | I'm trying to make sure I got the fpic thing right | 00:34 |
blackburn | ohh | 00:35 |
blackburn | bingo, I've got that too | 00:35 |
shogun-buildbot | build #552 of deb3 - modular_interfaces is complete: Failure [failed test python_modular] Build details are at http://www.shogun-toolbox.org/buildbot/builders/deb3%20-%20modular_interfaces/builds/552 blamelist: Sergey Lisitsyn <lisitsyn.s.o@gmail.com> | 00:35 |
adoniscik | We have a bingo!? | 00:35 |
blackburn | adoniscik: enable shared is not the thing | 00:35 |
blackburn | my mistake | 00:35 |
adoniscik | We have a bingo!? make CFLAGS=-fPIC CXXFLAGS=-fPIC | 00:36 |
adoniscik | http://www.mail-archive.com/nlopt-discuss@ab-initio.mit.edu/msg00018.html | 00:36 |
blackburn | yeah, that's the thing we should try | 00:37 |
blackburn | argh | 00:41 |
blackburn | I remember there was a more elegant way | 00:42 |
adoniscik | I couldn't get it to work; did you? | 00:43 |
blackburn | not yet | 00:44 |
blackburn | adoniscik: done | 00:49 |
adoniscik | oh yeah | 00:49 |
blackburn | okay what I did | 00:49 |
blackburn | CFLAGS="-fPIC" CXXFLAGS="-fPIC" ./configure --enable-shared | 00:49 |
blackburn | but do make clean before calling make | 00:49 |
adoniscik | oh so you set env vars | 00:50 |
blackburn | yeah | 00:50 |
blackburn | adoniscik: there is other way actually | 00:50 |
blackburn | a better way I'd say | 00:50 |
blackburn | ./configure --enable-shared | 00:50 |
blackburn | then open Makefile | 00:51 |
blackburn | find CXXFLAGS | 00:51 |
blackburn | and CFLAGS | 00:51 |
blackburn | and add -fPIC there | 00:51 |
blackburn | you may also change -O2 to -O3 to optimize it more :) | 00:51 |
adoniscik | then run configure --enable-shared ? | 00:51 |
adoniscik | done editing, what next? | 00:51 |
blackburn | make | 00:52 |
blackburn | and sudo make install | 00:52 |
blackburn | argh | 00:52 |
blackburn | make clean | 00:52 |
blackburn | make | 00:52 |
blackburn | sudo make install | 00:52 |
blackburn | without clean it could fail | 00:52 |
adoniscik | three commands for nlopt right? | 00:52 |
adoniscik | make commands that is | 00:53 |
blackburn | yes | 00:53 |
blackburn | and then try to 'make' shogun again | 00:53 |
blackburn | should work | 00:53 |
adoniscik | how did you multithread compile again? | 00:53 |
blackburn | make -jX | 00:54 |
blackburn | X is a number of threads | 00:54 |
adoniscik | no install? | 00:54 |
blackburn | install just installs so no need to multithread that | 00:55 |
blackburn | or what do you mean? | 00:55 |
adoniscik | i should run make not make install? | 00:55 |
adoniscik | for shogun | 00:55 |
blackburn | for shogun? | 00:55 |
blackburn | well just run make to check if it compiles | 00:55 |
blackburn | later you can sudo make install without recompiling it again | 00:55 |
blackburn | it recompiles only if you changed headers | 00:56 |
blackburn | okay it should work now :) | 00:58 |
blackburn | heiko: see you (again) :) take care | 00:59 |
adoniscik | it's been compiling for a while :) thanks | 00:59 |
adoniscik | and no | 00:59 |
adoniscik | interesting | 00:59 |
blackburn | adoniscik: heh again? | 00:59 |
adoniscik | no different | 00:59 |
heiko | blackburn, bye :) | 00:59 |
adoniscik | got the opld python problems again | 00:59 |
adoniscik | I thought we took care of that. maybe another round of configure is in order | 00:59 |
blackburn | adoniscik: argh.. anyway you know what to do, right? | 01:00 |
adoniscik | I'll try to take care of it; you helped plenty | 01:00 |
adoniscik | I am grateful | 01:00 |
blackburn | I'll be online next morning so we could talk about that | 01:00 |
blackburn | see you | 01:00 |
adoniscik | ok | 01:00 |
-!- blackburn [~blackburn@188.168.5.67] has quit [Quit: Leaving.] | 01:00 | |
-!- heiko [~heiko@host86-183-74-52.range86-183.btcentralplus.com] has left #shogun [] | 02:31 | |
shogun-buildbot | build #122 of nightly_default is complete: Failure [failed test] Build details are at http://www.shogun-toolbox.org/buildbot/builders/nightly_default/builds/122 | 03:45 |
adoniscik | anyone experts around? | 06:01 |
adoniscik | got a quick question about the Makefile | 06:02 |
Marty28 | any mkl-pros on? | 09:50 |
adoniscik | no, but I'm curious. what's the problem? | 09:51 |
Marty28 | I have some hundret scalars. | 09:51 |
Marty28 | Features | 09:51 |
Marty28 | floats | 09:51 |
Marty28 | I want to do MKL prediction on binary data | 09:52 |
Marty28 | I know I can use the Combined Kernel and MKL SVM. | 09:52 |
Marty28 | Still what kernel do I use for just 1 feature each? | 09:52 |
adoniscik | oh, that mkl. I thought you meant the Intel Math Kernel Library | 09:52 |
Marty28 | Gaussian works, but is it ideal for only 1D? | 09:52 |
Marty28 | haha | 09:53 |
Marty28 | no | 09:53 |
Marty28 | Could also have been ninja-related. | 09:53 |
Marty28 | http://software.intel.com/en-us/intel-mkl | 10:12 |
Marty28 | ah, cool | 10:13 |
adoniscik | :) | 10:15 |
adoniscik | which platform are you on? | 10:16 |
Marty28 | Intel, Ubuntu | 10:16 |
adoniscik | me too. which version? | 10:16 |
Marty28 | 64 bit | 10:16 |
adoniscik | of shogun | 10:16 |
Marty28 | 1.1.0 | 10:16 |
Marty28 | 2.0.0 is failty | 10:16 |
Marty28 | faulty | 10:16 |
adoniscik | what happens? | 10:17 |
Marty28 | Same code | 10:17 |
Marty28 | Same data | 10:17 |
Marty28 | 1.1.0 does Multiple Kernel Learning-MKL | 10:17 |
Marty28 | nice result | 10:17 |
Marty28 | 2.0.0 only crap | 10:17 |
adoniscik | did you report it? | 10:17 |
Marty28 | I hear about it often. I thought they knew it is instable. | 10:18 |
Marty28 | WeightedDegreePositionStringKernel | 10:18 |
Marty28 | 2 of them | 10:18 |
Marty28 | alone they perform | 10:18 |
Marty28 | combined they suck | 10:18 |
Marty28 | if 2.0.0 | 10:18 |
adoniscik | really? I didn't see the warnings | 10:18 |
adoniscik | but it is to be expected with a 0.0 release, I suppose. | 10:19 |
Marty28 | I guess they changed the interior of MKL. | 10:19 |
Marty28 | So it is not optimized for the legacy WeightedDegreePositionalStringKernel | 10:19 |
Marty28 | Also the MKL crashes if there is nothing to find | 10:20 |
adoniscik | you can examine the code if you are determined | 10:21 |
adoniscik | https://github.com/shogun-toolbox/shogun/tree/master/src/shogun/classifier/mkl | 10:21 |
Marty28 | i know | 10:22 |
adoniscik | which interfaces are you using? | 10:22 |
Marty28 | python_modular | 10:22 |
Marty28 | only | 10:22 |
adoniscik | I've been having trouble compiling that one in 2.0 | 10:22 |
adoniscik | which python distribution do you use? | 10:22 |
Marty28 | 2.7.3 | 10:22 |
adoniscik | from the ubuntu software center? | 10:23 |
Marty28 | python yes | 10:23 |
Marty28 | shogun i compiled | 10:23 |
adoniscik | I got Enthought's thinking it would help but it's not working with shogun yet | 10:23 |
adoniscik | http://www.enthought.com/products/epd.php | 10:23 |
adoniscik | did it work out of the box in 2.0 or did you have to modify any files? | 10:24 |
adoniscik | I know that you reverted to 1.1 | 10:24 |
Marty28 | there are parameters | 10:24 |
Marty28 | as far as i remember it worked just the same way that also worked in 1.1 | 10:25 |
Marty28 | i will check my logs | 10:25 |
adoniscik | thanks | 10:25 |
-!- blackburn [5bdfb203@gateway/web/freenode/ip.91.223.178.3] has joined #shogun | 10:26 | |
blackburn | hey there | 10:26 |
adoniscik | hey !! | 10:26 |
Marty28 | darn, how do i paste stuff here? | 10:26 |
adoniscik | good morning | 10:26 |
Marty28 | huho | 10:26 |
Marty28 | hiho | 10:26 |
blackburn | Marty28: what is faulty in 2.0? :) | 10:26 |
adoniscik | pastebin, marty | 10:26 |
Marty28 | http://pastebin.com/sRtrC0uX | 10:27 |
Marty28 | Multiple Kernel Learning does not work perfectly. | 10:28 |
Marty28 | In 1.1 the results are ok. | 10:28 |
blackburn | really? | 10:28 |
Marty28 | In 2.0 they are inexsitent. | 10:28 |
adoniscik | marty: well that's the configure, but what about the make? | 10:28 |
Marty28 | Depends on the Kernel, of course. | 10:28 |
blackburn | can you prepare some kind of snippet to reproduce that? | 10:29 |
blackburn | we'd appreciate | 10:29 |
adoniscik | blackburn: I have the -ldl -lutil issue during the python_modular stage of make but I couldn't figure how to pass the flags. I'm looking at src/interfaces/python_modular/Makefile. | 10:29 |
blackburn | adoniscik: no, all options are in src/Makefile | 10:30 |
blackburn | and in configure script | 10:30 |
Marty28 | CombinedKernel was filled with 2 x WeightedDegreePositionalStringKernel | 10:31 |
adoniscik | I can't find any interface names in the Makefile | 10:31 |
adoniscik | I get make[1]: Entering directory `/shogun/src/interfaces/python_modular' | 10:32 |
blackburn | adoniscik: what is the command that fails on compilation? | 10:32 |
adoniscik | c++ -g -fPIC -O0 -DSWIG_TYPE_TABLE=shogun -DSHOGUN -DLINUX -DUSE_BOOL -DUSE_CHAR -DUSE_UINT8 -DUSE_UINT16 -DUSE_INT32 -DUSE_INT64 -DUSE_UINT64 -DUSE_FLOAT32 -DUSE_FLOAT64 -DHAVE_POWL -DHAVE_LGAMMAL -DHAVE_SQRTL -DHAVE_LOG2 -DHAVE_PTHREAD -DHAVE_HDF5 -DHAVE_XML -DHAVE_ATLAS -DHAVE_LAPACK -DHAVE_SUPERLU -DHAVE_ARPACK -DHAVE_NLOPT -DHAVE_EIGEN3 -DUSE_GLPK -DUSE_LZO -DUSE_GZIP -DUSE_BZIP2 -DUSE_LZMA -DHAVE_LARGEFILE -DUSE_SHORTREAL_KERNE | 10:32 |
adoniscik | LCACHE -DUSE_BIGSTATES -DUSE_HMMCACHE -DUSE_REFERENCE_COUNTING -DUSE_SVMLIGHT -DUSE_SPINLOCKS -DHAVE_PYTHON -c -I. -I.. -I/usr/include/libxml2 -I/usr/include/atlas -isystem /usr/include/eigen3 -I../../shogun -I../.. -I/opt/enthought/include/python2.7 -I/opt/enthought/lib/python2.7/site-packages/numpy/core/include -o modshogun_wrap.cxx.o modshogun_wrap.cxx | 10:32 |
adoniscik | c++ -Xlinker --no-undefined -L../../shogun -lshogun -ldl -lutil -shared -L/opt/enthought/lib/python2.7/config -lpython2.7 -ldl -lutil -o _modshogun.so modshogun_wrap.cxx.o sg_print_functions.cpp.o -L../../shogun -lshogun -lm -pthread -lhdf5 -lxml2 -llapack -lcblas -llapack_atlas -lsuperlu -larpack -lnlopt -lglpk -lm -llzo2 -lz -lbz2 -llzma -shared -L/opt/enthought/lib/python2.7/config -lpython2.7 | 10:32 |
adoniscik | I think I need to append -lutil -ldl but I don't know how | 10:33 |
adoniscik | it already has it in several places, but maybe the wrong ones? | 10:33 |
Marty28 | mom | 10:33 |
Marty28 | long ago | 10:34 |
blackburn | yeah already have it twice | 10:37 |
blackburn | but still the error is there, right? | 10:37 |
adoniscik | yes, the one we talked about before | 10:37 |
adoniscik | let me try to run that command myself from the CLI | 10:38 |
adoniscik | I just noticed I pasted two c++ commands; the first one obviously worked | 10:40 |
adoniscik | it compiles fine when you move -ldl -lutil to the end; I delete the extraneous earlier ones | 10:41 |
adoniscik | so the question is how to get this in the Makefile | 10:42 |
blackburn | adoniscik: hmm I see | 10:43 |
blackburn | adoniscik: I am currently have to get some things done at job | 10:43 |
adoniscik | I actually don't need it right now since doing that one line manually and re-executing make successfully completed the run, but it should be resolved for a future release... | 10:43 |
blackburn | yeah | 10:43 |
blackburn | I'll get back to that later today | 10:44 |
adoniscik | thanks. I (probably) have a working installation now. | 10:46 |
Marty28 | super | 10:46 |
adoniscik | what's the latest versionh. the docs reference 2.0.1 but the source says 2.0.0 | 10:46 |
blackburn | adoniscik: 2.0.1 is to be released | 10:48 |
adoniscik | is it normal for make to return this warning? warning: ‘int _import_array()’ defined but not used [-Wunused-function] | 10:50 |
blackburn | adoniscik: yes | 10:51 |
adoniscik | blackburn: thanks. I did a make install and ran shogun now. I guess I should run the tests for good measure. | 10:53 |
-!- adoniscik [~emre@c-67-180-103-118.hsd1.ca.comcast.net] has quit [Ping timeout: 256 seconds] | 11:38 | |
blackburn | Marty28: are you using MKL for regression? | 12:11 |
Marty28 | no | 12:15 |
Marty28 | binary | 12:15 |
Marty28 | 2-class | 12:15 |
Marty28 | Yesterday it crashed when it could not find anything, unfortunately I cant reproduce it. | 12:16 |
Marty28 | Or fortunately. | 12:16 |
Marty28 | That was with Gaussian Kernels. | 12:16 |
blackburn | I remember we had some problem with LINADD optimization | 12:20 |
Marty28 | At the moment I have some 100 of real features and I want an MKL with importances. | 12:25 |
Marty28 | I do MKL with e.g. 100 features, each separately in its own subkernel. | 12:26 |
Marty28 | Is the GaussianKernel a good idea for the subkernel if it is only 1-D? | 12:26 |
Marty28 | It is working for toy data, but I need it for reallife data. | 12:27 |
blackburn | you do feature selection using mkl?? | 12:31 |
Marty28 | I seek feature importance. | 12:32 |
Marty28 | I have n DNA regions. | 12:32 |
blackburn | well you could do that using L1 regularization as well | 12:32 |
Marty28 | From there I derive m features each. | 12:32 |
blackburn | if you have group sparsity you could use FeatureBlockLogisticRegression I wrote before | 12:33 |
Marty28 | Then I do MKL with n * m GaussianKernels via MKL. | 12:33 |
Marty28 | Feature block sounds interesting. | 12:33 |
blackburn | it allows group L1/Lq and tree L1/L2 regularization | 12:34 |
Marty28 | I am a biologist, now bioinformaticist. | 12:34 |
blackburn | like (L1 of [ (L2 of group 1) (L2 of group 2) .. () ] ) | 12:34 |
Marty28 | So what is it good for? *g* | 12:34 |
Marty28 | What is regularization? | 12:35 |
blackburn | Marty28: well for example I applied it for metagenomics where various species are of tree structure | 12:35 |
blackburn | Marty28: general learning framework is loss + regularization | 12:35 |
blackburn | loss imposes constraint on how do you measure error on a train set | 12:36 |
blackburn | regularization is like a prior information (in SVM it is a L2 regularization used) | 12:36 |
blackburn | in case of mixed (like L1/Lq) regularization you pull weight of each group to be similar | 12:37 |
blackburn | but in the same time you pull irrelevant groups of features to have zero norm which means they all become zero | 12:38 |
Marty28 | I guess I have to look that up. Is it normal content in Machine Learning books? | 12:38 |
blackburn | loss + regularization is everywhere, group sparsity things are not that popular I think | 12:38 |
Marty28 | fine | 12:39 |
Marty28 | I want to predict my binary classes and see, which group (sequence region) and which features there were important. | 12:40 |
blackburn | Marty28: I could reference http://arxiv.org/pdf/0909.1373.pdf | 12:40 |
blackburn | Marty28: group sparsity learning does that - with it you are trying to find a best classifier with given regularization | 12:40 |
blackburn | the more powerful regularization is (like bigger lambda coefficient) - the more groups go out | 12:41 |
blackburn | so say you have 3 groups with 3 features in each of them then you may get something like w = [0,0,0,0.13,0.3,-0.1,0,0,0] | 12:42 |
Marty28 | I want to develop features from DNA-sequences that are helpful in MKL | 12:42 |
blackburn | which means group 1 and group 3 were less relevant | 12:42 |
blackburn | and feature 2 of group 2 was the most relevant | 12:42 |
blackburn | see what I mean? | 12:43 |
Marty28 | yes | 12:43 |
Marty28 | is what i did | 12:43 |
Marty28 | I can get negative weights? | 12:43 |
blackburn | yes, in that framework you can | 12:44 |
Marty28 | What is the relevance of the blocks? | 12:44 |
blackburn | it is a linear model ( f(x) = sign(<w,x> + b)) vector | 12:44 |
blackburn | relevance of block is a L2 norm of it | 12:45 |
Marty28 | Is it just coefficients for adding up the kernels? | 12:45 |
Marty28 | (the weights) | 12:45 |
blackburn | it is not for kernels - only for features | 12:45 |
blackburn | if you have to stay with kernels MKL is the only way I think | 12:46 |
Marty28 | I just derive features and then I want to know who was helpful and from which DNA region it came. | 12:46 |
blackburn | but if you have real features -- group sparsity algorithms would work too | 12:46 |
blackburn | Marty28: are your features dense and real? | 12:47 |
Marty28 | real numbers | 12:47 |
Marty28 | What does dense mean? | 12:47 |
blackburn | not sparse (like >50% of non -zeroes) | 12:47 |
Marty28 | ah | 12:47 |
Marty28 | yes, most of them dense | 12:47 |
Marty28 | ok, it depends | 12:48 |
blackburn | you could try that feature block thing then | 12:48 |
Marty28 | might get sparse through cutoffs. | 12:48 |
blackburn | it is pretty easy | 12:48 |
blackburn | so you have RealFeatures already, right? | 12:48 |
Marty28 | I am generating them, yes. | 12:48 |
blackburn | I am talking about code now | 12:48 |
blackburn | okay | 12:48 |
blackburn | all you have to do is | 12:48 |
Marty28 | ok | 12:48 |
blackburn | create IndexBlockGroup and pass features,labels and the group to FeatureBlockLogisticRegression | 12:49 |
blackburn | Marty28: see the example https://github.com/shogun-toolbox/shogun/blob/master/examples/undocumented/python_modular/classifier_featureblock_logistic_regression.py | 12:49 |
blackburn | Marty28: it forms two groups of features | 12:50 |
blackburn | like first half and second half | 12:51 |
Marty28 | nice | 12:51 |
blackburn | 0.1 in FeatureBlockLogisticRegression stands for regularization parameter | 12:51 |
Marty28 | For me it will be e.g. 10 DNA regions | 12:51 |
blackburn | when set_regularization(1) that parameter should be in [0,1] range | 12:52 |
blackburn | 1 means all zeros | 12:52 |
blackburn | 0 means no regularization (all are not zeros) | 12:52 |
blackburn | you would have to add 10 IndexBlocks then | 12:52 |
blackburn | containing (first index of feature of the DNA region, last_index of feature of the DNA region+1) | 12:53 |
blackburn | I am leaving for a lunch now, be back in 20 mins | 12:54 |
Marty28 | ok | 12:55 |
Marty28 | good idea | 12:55 |
Marty28 | thanks | 12:55 |
Marty28 | What if e.g. in Feature7 I have 2 regions with label +1 and the rest is label -1? | 13:15 |
Marty28 | meaning: - - - - + + + - - - - - - - + + - - | 13:17 |
Marty28 | And it is not separable through other features | 13:17 |
Marty28 | GaussianKernel does it | 13:19 |
Marty28 | ok, depends on the regression | 13:21 |
blackburn | back | 13:32 |
blackburn | Marty28: I am not sure I've got what you mean | 13:32 |
blackburn | if you are talking about labels it is a different 'axis' | 13:34 |
Marty28 | I mean: At positive examples (label +1) feature 7 is most times around 3 OR around 7 | 14:00 |
Marty28 | So there are 2 maxima in the regression curve, I guess. | 14:02 |
blackburn | ahh | 14:43 |
-!- heiko [~heiko@host86-183-74-52.range86-183.btcentralplus.com] has joined #shogun | 14:49 | |
blackburn | heiko: how one does make verbs to be in past in German? :) | 15:08 |
heiko | no idea | 15:08 |
blackburn | I was so tired of lecture yesterday that learned a little deutsch | 15:08 |
heiko | There is a rule probably | 15:08 |
blackburn | how could I say 'I came'? | 15:08 |
heiko | but I only know forms | 15:08 |
heiko | ich kam | 15:09 |
blackburn | kam? | 15:09 |
blackburn | oh | 15:09 |
heiko | gehen - ich bing | 15:09 |
heiko | ging | 15:09 |
heiko | programmieren - ich programmierte | 15:09 |
blackburn | heiko: is there any perfect forms? | 15:09 |
blackburn | I have done | 15:09 |
blackburn | like ^ | 15:09 |
heiko | yeah | 15:11 |
heiko | ich bin gegangen | 15:11 |
heiko | ich habe programmiert | 15:11 |
blackburn | hehe | 15:13 |
blackburn | heiko: so basically all the rules are nearly the same? | 15:13 |
heiko | to be honest: I have no idea :D | 15:13 |
blackburn | heiko: btw do you learn a lot of grammar things at school? | 15:13 |
heiko | I usually dont learn languages like this, in particular not my native | 15:13 |
heiko | I like to read/talk | 15:14 |
blackburn | it is the only way to learn new language | 15:14 |
heiko | yeah I learned this stuff, but i was bad in it and also thats ten years ago :D | 15:14 |
blackburn | that's a problem of russian english | 15:14 |
blackburn | we all study a lot of gerunds, infinitives, etc | 15:14 |
blackburn | but all russians are bad in english | 15:14 |
blackburn | when I was 2nd year undergraduate I studied functions of gerund in english :D | 15:15 |
blackburn | no idea what for | 15:15 |
heiko | lol | 15:16 |
heiko | talking is the best | 15:16 |
heiko | thing you can do | 15:16 |
heiko | and reading | 15:16 |
heiko | newspaper or books | 15:16 |
heiko | not chat :) | 15:16 |
blackburn | well I have no one to talk here :) | 15:16 |
heiko | I think there are also online teachers who talk to you via skype | 15:17 |
blackburn | ah well we've got a lot of guests from africa in dormitories | 15:17 |
blackburn | but it is anti-learning | 15:17 |
heiko | better speak real english :) | 15:17 |
heiko | african english is hard, same as african french | 15:18 |
blackburn | I have no idea what do they speak about | 15:18 |
blackburn | som bububububu | 15:18 |
blackburn | some* | 15:18 |
blackburn | heiko: ha! I recalled | 15:18 |
blackburn | when I was 13 I've been studying german at school | 15:19 |
blackburn | I had to remember what is Partizip Zwei | 15:19 |
blackburn | not how to talk, how to read | 15:19 |
blackburn | PARTIZIP ZWEI | 15:19 |
blackburn | :D | 15:19 |
blackburn | zero chance to learn anything right? :) | 15:19 |
heiko | I dont even kno what that is :) | 15:19 |
heiko | I always had the worst marks in coma exams :) | 15:20 |
blackburn | heiko: like 'keys that were lost' | 15:21 |
heiko | I see | 15:21 |
blackburn | some characteristic of an object or so | 15:21 |
blackburn | okay time to start early weekend | 15:22 |
blackburn | heiko: see you | 15:22 |
heiko | wow thats early :) | 15:23 |
heiko | see you blackburn, enjoy! | 15:23 |
-!- blackburn [5bdfb203@gateway/web/freenode/ip.91.223.178.3] has quit [Quit: Page closed] | 15:23 | |
Marty28 | Feature request: Some kinds of progress bars everywhere... | 15:59 |
Marty28 | -s | 15:59 |
heiko | Marty28, what exactly do you mean? which interface are you using=? | 16:03 |
Marty28 | python_modular | 16:08 |
Marty28 | MKl takes a long time | 16:08 |
Marty28 | With only 20 features and crappy data. | 16:08 |
Marty28 | It is fast when it is easy ;) | 16:09 |
Marty28 | Does it have to converge so you do cannot estimate how long the training will last? | 16:10 |
heiko | yes | 16:11 |
heiko | Marty28, you can always turn on debug or information messages | 16:12 |
heiko | The iteration numbers are then displayed | 16:12 |
Marty28 | how do i do this? | 16:12 |
heiko | let me check, I always forget how to do this for python .. 1min | 16:13 |
Marty28 | Yesterday with nothing to find it just crashed. | 16:13 |
Marty28 | thx | 16:13 |
heiko | OBJECT.io.set_loglevel(MSG_DEBUG) | 16:14 |
heiko | and you will have to import MSG_DEBUG | 16:14 |
heiko | from shogun.Classifier import MSG_DEBUG | 16:15 |
heiko | if you want to debug a classifier | 16:15 |
heiko | if you give precise problem descriptions, we can fix the bug | 16:15 |
Marty28 | cool | 16:19 |
Marty28 | thx | 16:19 |
Marty28 | works fine | 16:21 |
-!- K0stIa [~kostia@2001:718:2:1634:4a5b:39ff:fe99:cc49] has joined #shogun | 16:42 | |
K0stIa | Hi, all. I would like to try shogun's BMRM implementation, but I didn't figure out if it is possible to define own first-order oracle in python and give it through input parameters (like a class or function) to BMRM procedure. Could someone answer me if it is possible to define in python own oracle? | 16:47 |
-!- n4nd0 [~nando@n188-p189.kthopen.kth.se] has joined #shogun | 16:50 | |
heiko | K0stIa, hi the BMRM is pretty new. | 17:00 |
K0stIa | heiko: Hi. | 17:01 |
heiko | we recently added Director classes which is the possibility to prototype in python and then use that code | 17:01 |
heiko | but thats also new | 17:01 |
K0stIa | heiko: I know it;s appeared recently | 17:01 |
heiko | Since I did not implement that, I guess it is best that you write to the mailing list and provide an accurate description of what you want to do | 17:01 |
heiko | I think we can do this somehow, but dont know how currently | 17:02 |
heiko | its similar to defining an own kernel for svm (which is possible in python) | 17:02 |
K0stIa | heiko: but such thing would be really cool and could increase number of shogun users... | 17:03 |
heiko | I agree :) | 17:03 |
heiko | I currently have some time this week -- if you write some more details, I might be able to help there, but I will have to get back to bmrm and Director classes people | 17:03 |
K0stIa | so, I didn't get what is the current situation ? I understood it like there is some implementation but it is not tested, or what ? | 17:04 |
heiko | no the bmrm stuff should be working, it is part of the latest release | 17:04 |
K0stIa | I can define my own oracle in python and give it as input parameter to BMRM ? | 17:05 |
heiko | that not I meant the BMRM itself | 17:06 |
heiko | prototyping in python is new | 17:06 |
K0stIa | yep, it could be done in c++. | 17:06 |
heiko | I think so | 17:06 |
K0stIa | but my question if it is possible to do in python ? | 17:06 |
heiko | I would say yes, but this would have to be added manually, which is not too hard as far as I remember | 17:07 |
heiko | therefore: ask mailinglist | 17:07 |
heiko | gsomix did the Director Classes that allow such things | 17:07 |
K0stIa | can you give me his contacts ? email, for instance | 17:08 |
heiko | you should write to the list, but I can also give you their private addresses, wait | 17:09 |
heiko | Evgeniy Andreev gsomix@gmail.com | 17:09 |
heiko | (director classes) | 17:09 |
heiko | Michal U?i??? uricar.michal@gmail.com (wrote the bmrm stuff) | 17:10 |
heiko | if you write to the list, its more probably that you will get help since other developers also see it | 17:10 |
K0stIa | ok, I will try to contact Evgeniy Andreev | 17:10 |
K0stIa | I saw his name in swig configuration files... | 17:11 |
K0stIa | he could help me | 17:11 |
heiko | As said: he did a similar thing for kernels | 17:12 |
heiko | gotta go, see you later! | 17:13 |
K0stIa | see you. | 17:14 |
K0stIa | thanks | 17:14 |
-!- heiko [~heiko@host86-183-74-52.range86-183.btcentralplus.com] has quit [Quit: Leaving.] | 17:15 | |
-!- n4nd0 [~nando@n188-p189.kthopen.kth.se] has quit [Quit: leaving] | 18:04 | |
-!- K0stIa [~kostia@2001:718:2:1634:4a5b:39ff:fe99:cc49] has left #shogun [] | 18:11 | |
-!- heiko [~heiko@host86-183-74-52.range86-183.btcentralplus.com] has joined #shogun | 18:17 | |
-!- adoniscik [~emre@c-67-180-103-118.hsd1.ca.comcast.net] has joined #shogun | 19:04 | |
-!- heiko [~heiko@host86-183-74-52.range86-183.btcentralplus.com] has quit [Quit: Leaving.] | 19:25 | |
-!- blackburn [~blackburn@31.28.33.88] has joined #shogun | 19:45 | |
-!- heiko [~heiko@host86-183-74-52.range86-183.btcentralplus.com] has joined #shogun | 20:11 | |
-!- heiko [~heiko@host86-183-74-52.range86-183.btcentralplus.com] has quit [Client Quit] | 20:13 | |
-!- Marty28 [9eb54d46@gateway/web/freenode/ip.158.181.77.70] has quit [Quit: Page closed] | 20:13 | |
-!- zxtx [~zv@2602:306:32c1:5960:f489:6a3b:a73b:c233] has joined #shogun | 20:35 | |
-!- zxtx [~zv@2602:306:32c1:5960:f489:6a3b:a73b:c233] has quit [Ping timeout: 246 seconds] | 22:31 | |
adoniscik | more news to report. first, I can't get shogun to load in EPD; it can't find it. Second, I get crashes in the native python. | 23:22 |
blackburn | adoniscik: amazing news :) | 23:23 |
blackburn | I'd like more details | 23:23 |
adoniscik | yeah, really! trying to get a list of installed modules does this: http://pastebin.com/KvF6EUch | 23:23 |
blackburn | funtastic | 23:24 |
adoniscik | it's launched a bug report prompt pointing to _modshogun.so, libshogun.so.12.0 and libnlopt.so.0.7.0 adding "It is highly recommended to check if the problem persists without those first." | 23:25 |
blackburn | adoniscik: I have no idea about that.. | 23:33 |
blackburn | apparently EPD is so troublesome | 23:35 |
adoniscik | the pastebin is from the regular one; I can't get it to work with EPD at all yet | 23:35 |
adoniscik | should I add some folders to PYTHONPATH to help EPD find shogun? | 23:39 |
blackburn | adoniscik: I am a bit lost with these errors - try if you think that could help | 23:40 |
adoniscik | okay. where does shogun install its modules? | 23:41 |
blackburn | .so? | 23:41 |
adoniscik | the ones that are called with the "import" command in Python | 23:41 |
adoniscik | "import shogun" | 23:41 |
blackburn | let me recall | 23:41 |
blackburn | /usr/local/lib/python2.7/dist-packages/ | 23:42 |
blackburn | on my machine | 23:42 |
blackburn | something like that on yours as well I guess | 23:42 |
adoniscik | good, you are right | 23:42 |
adoniscik | it has four files | 23:42 |
--- Log closed Fri Oct 05 00:00:17 2012 |
Generated by irclog2html.py 2.10.0 by Marius Gedminas - find it at mg.pov.lt!