Open in new window / Try shogun cloud
--- Log opened Mon Feb 06 00:00:13 2017
@sukeyPull Request #3608 "Clean up KNN"  synchronized by MikeLing -
@sukeyPull Request #3584 "LinalgRefactor - Apply"  synchronized by OXPHOS -
-!- mikeling [uid89706@gateway/web/] has joined #shogun08:28
-!- praisethemoon [~praisethe@] has joined #shogun10:37
-!- praisethemoon [~praisethe@] has quit [Ping timeout: 248 seconds]11:12
-!- praisethemoon [~praisethe@] has joined #shogun12:01
-!- praisethemoon [~praisethe@] has quit [Changing host]12:02
-!- praisethemoon [~praisethe@unaffiliated/praisethemoon] has joined #shogun12:02
-!- HeikoS [] has joined #shogun12:05
-!- mode/#shogun [+o HeikoS] by ChanServ12:05
-!- ShellcatZero [] has quit [Ping timeout: 245 seconds]12:07
@wikinglisitsyn, pppppppppuuuung12:30
lisitsynwiking: peng12:35
@wikingorg.shogun-toolbox.* or ml.shogun.*12:45
@wikingas java pkgs?12:45
-!- praisethemoon [~praisethe@unaffiliated/praisethemoon] has quit [Ping timeout: 248 seconds]12:56
CaBawiking: heya12:57
CaBaHeikoS: hey, you happen to be around?12:58
@HeikoSCaBa: hi yes12:58
CaBaHeikoS: remember when we were talking the other day about kernel functions that depend not only on something that cannot be plugged into a CFeature class, but for example index structures and things to speed up the computation of K(x_1, x_2)? And you said one solution would be to precalc the kernel matrix and use CCustomKernel but that that would not support e.g. normalizers?13:00
@HeikoSi remember13:00
@wikingCaBa, yes13:00
@HeikoSwiking:  my plan for today: write GSoc projects13:01
@HeikoSwiking: you got some bullet points on that?13:01
@HeikoSsame projects as last year more or less? ;)13:01
CaBaHeikoS: i know this is dirty, but would this work: use CDenseFeatures<size_t>, 1-dimensional, the size_t value actually being an index for a feature vector... then derive from CKernel, override ::compute(), which then performs lookups according to the size_t feature in the "real" kernel matrix and returns the value. that would be a fully functional kernel, wouldnt it? one that supports subsetting and therefor13:04
CaBae x-val, normalization during x-val, MKL, etc? no restrictions?13:04
@HeikoSyes that would work13:06
@HeikoSbut it is a hack13:06
@HeikoSas your kernel would only work with this very specialized features13:06
@HeikoSgenerally, if you implement subsetting on kernel and features, then you can do anything you like and it will work13:07
@HeikoS"anything you like" as in override methods in a weird way13:07
CaBaHeikoS: well the features wouldn't be visible to the user - just like in CCustomKernel. The kernel would just take a precomputed matrix in it's constructor.13:07
CaBaHeikoS: it would be kind of a hacky replacement for CCustomKernel that supports normalization13:08
@HeikoSwhat I mean is that you have the inheritance diagram, and there two leafs (i.e. specialized classes) depend on one another13:08
@HeikoSwhich is against OOP paradigma13:08
@HeikoSbut yeah13:08
@HeikoSI see your need for hacks ;)13:08
@HeikoSwould be a cool thing to do as part of a GSoC project on x-validation and modelselection13:09
CaBamy need for hacks? ^^13:09
@HeikoSyou seem to have this complicated kernel13:10
@HeikoSbtw did you ever think about just not normalising the x-validation kernel matrices?13:10
@HeikoSi.e. just normalise your precomputed kernel matrix?13:10
CaBasure, but it's a different thing13:10
@HeikoSor is it really crucial?13:11
CaBai don't know if it's relevant13:11
@HeikoSI could also btw imagine some on-the-fly normalisation that works with custom kernels13:11
@wikingHeikoS, ah man13:11
@wikingi have to remember all those good ideas13:11
@wikingwe have in the parking lot13:11
@HeikoSwiking: yes that is what I am asking :)13:11
@HeikoSwiking: I have a few things13:11
@HeikoSbut before I tell them, I wanted to ask to get more coverage13:11
@wikingHeikoS, btw we need to rename our java package... it's either org.shogun-toolbox.* or ml.shogun.*13:12
@wikingas we do not own org.shogun13:12
@wikingi mean shogun.org13:12
@HeikoSuh, nice13:12
CaBabtw - do you know whether per-kernel sqrt-diag normalization is sufficient in MKL to get interpretable subkernel weights? or if other means of normalization are necessary?13:12
@wikingHeikoS, so we need to make a decision here...13:12
@HeikoSCaBa: for xvalidation, you dont need interpretable weights or do you?13:12
@HeikoSwiking: ml.shogun13:13
@HeikoSCaBa: or just with the final model13:13
@wikingi'll register13:13
@wikinglisitsyn, ^13:13
@HeikoSwiking: cool!13:13
@HeikoSCaBa: because if you just tune parameters using xvalidation, then maybe it is not that important13:13
CaBaHeikoS: that's what i'm wondering - will i just get different betas if i normalize afterwards but the subkernel parameters remain optimal?13:14
@HeikoSCaBa: probably it is worth a try13:14
@HeikoSbut I dont know much about MKL13:15
@HeikoSand sonney2k who is the only one who knows, is busy with life13:15
CaBaHeikoS: life is a cool thing to be busy with ;)13:18
@HeikoSCaBa: so just to get you right13:18
@HeikoSCaBa: you want13:18
@HeikoSto use13:18
@HeikoScustom kernel + mkl + cross-validation + kernel normaliser13:19
@HeikoSmaybe we can just make this an entrance task13:19
@HeikoSwe got some students incoming for GSoC13:19
@HeikoSand this might be a nice non-trivial thing to work out13:19
CaBathat sounds about right13:19
@HeikoSso what we need here is that the kernel normalisation works with the subsetting and the mkl, all on the fly13:20
CaBawell... they hacky replacement for CCustomKernel I had in mind is a 30 min job I guess. But I'm sure you had something clean in mind :D13:20
@HeikoSyes sure, I mean for your purposes, the 30min hack might be ok13:20
@HeikoSbut It is good for us to get such things done properly, in case somebody else also wants to use this kind of thing13:21
CaBaThe only ting CCustomKernel lacks in that respect might be the normalization. Last time we spoke you said x-val and mkl might just work.13:21
CaBaBut I didn't check that I must admit13:21
lisitsynwiking: ml.shogun if we don't lost it again13:24
@wikinglisitsyn, i've bought it for 5 years :)13:25
lisitsynorg.shogun-toolbox is impossible13:25
lisitsynwould be13:25
CaBaSG_REF checks for nullptr, right?13:31
@HeikoSbut best read the def13:32
CaBawhat's that deprecated info in the wiki? is there a replacement?13:32
@HeikoSno :)13:33
@sukeyPull Request #3608 "Clean up KNN"  synchronized by MikeLing -
mikelinghi, i got a bunch of undocument warning and failed to build python modular
-!- praisethemoon [~praisethe@] has joined #shogun14:56
mikelingwiking: ping15:20
mikelingwiking: after I pull docker image, how could I use python interface?15:22
mikelingon the image i mean15:23
@wiking - docker run -t -d -P -e "JAVA_HOME=/usr/lib/jvm/java-8-oracle" -e "CC=$CC" -e "CXX=$CXX" --name devenv -v $PWD:/opt/shogun shogun/shogun-dev /bin/sh -c "mkdir /opt/shogun/build;bash"15:23
@wiking - docker exec -t devenv /bin/sh -c "cd /opt/shogun/build; cmake -DCMAKE_INSTALL_PREFIX=$HOME/shogun-build -DENABLE_TESTING=ON $CMAKE_OPTIONS .."15:23
mikelingand pip install shogun-ml doesn't work, it looks like there is no package named  shogun-ml?15:24
@wikingpip install?15:26
@wikingwhy do you want pip install when you try a local development15:26
mikelingi just see that on document and want to have a try15:26
@wikingah forget it15:27
@wikingit needs further fixing15:27
@wikinghopefully 5.1.0 will have it15:27
mikelingok ;)15:27
mikelingwiking: It says "CMake Error: The source directory "/opt/shogun" does not appear to contain CMakeLists.txt. Specify --help for usage, or press the help button on the CMake GUI."15:50
@wikinglook i'll write this the one time15:50
@wikingdocker run -t -i -v <the path to your checked out source>:/opt/shogun shogun/shogun-dev /bin/bash15:51
@wiking(now you are in docker)15:51
@wikingcd /opt/shogun15:51
@wikingmkdir build15:51
@wikingcd build15:51
@wikingcmake .... (all the stuff above)15:51
@HeikoSmikeling: is there anything that is not documented?15:54
mikelingHeikoS: yep, part of warning is in
@HeikoSI mean as in documented installation process on our website15:56
@HeikoSthese are just warnings15:56
mikelingHeikoS: oh, the pip install shogun-ml doesn't work15:56
mikelingit seems like shogun package is missing15:56
@HeikoSmikeling: I think I dont understands what you are actually doing there. these were just 3 very unrelated things (docker, pip install, cmake)15:59
mikelingHeikoS: because I would like to install the python interface of shogun, after I failed to build it from source code locally, I switch to the pip install but it looks like missing shogun-ml package for it16:01
mikelingand now, I want to use docker to solver my problem16:01
mikelingBTW, the error message is /bin/sh: line 1: 13156 Segmentation fault: 11  /usr/local/bin/doxygen modshogun.doxy when I build python interface locally16:04
-!- HeikoS [] has quit [Ping timeout: 240 seconds]16:05
mikelingAnd here is the Cmake I used " cmake -DPYTHON_INCLUDE_DIR=/usr/local/Cellar/python/2.7.13/Frameworks/Python.framework/Headers  -DPYTHON_LIBRARY=/usr/local/Cellar/python/2.7.13/Frameworks/Python.framework/Versions/2.7/lib/libpython2.7.dylib  -DPythonModular=ON ../"16:05
@wikingmikeling, ?16:08
@wikingi told you how you can do it with the dokcer16:08
@wikingbut you can do the same way on your osx16:08
@wikingcmake -DPythonModular=ON  ..16:09
@wikingeither way works16:09
mikelingwiking: yeah, I can do the same thing on osx, but it doesn't work for me, at least it failed on locally without docker16:12
@wikingwhat's the failure?16:12
@wikingbtw without your patches16:12
@wikingor with your patches?16:12
mikelingwithout my pr16:13
mikelings /bin/sh: line 1: 13156 Segmentation fault: 11  /usr/local/bin/doxygen modshogun.doxy16:13
mikelingwithout the "s", sorry16:13
@wikingbrew info doxygen16:13
mikelingdoxygen: stable 1.8.13 (bottled), HEAD16:14
@wikinggit git rev-parse HEAD16:16
@wikingbut in worst case you can disable doxygen and do it without that16:17
mikelingyou mean git rev-parse HEAD?16:17
mikelingit's 9673755bdb345b3c10fe447425394e4813f06dd816:18
@wikingmmm i've just ran on my osx16:18
@wikingand i dont get segfault16:18
mikelingand here is my cmake output
@wiking-- Found NumPy: version "1.12.0" /usr/local/Cellar/matplotlib/2.0.0/libexec/lib/python2.7/site-packages/numpy/core/include?16:18
@wikingthis doesn't look too good16:19
mikelinghmmm, so you think it may case a segfault?16:20
@wikingbu ti dont understand16:20
@wikingwhy the numpy is part of matplotlib16:20
@wikingpip list|grep numpy16:20
mikelingit return numpy (1.12.0)16:20
@wikingpip show numpy16:21
mikelingdo you think docker can help me solve that segfault?16:22
@wikingwith docker you would get exactly the same thing16:22
@wikingwhat happens on travis16:22
@wikingas it's the same env16:22
@wikingyeah i have no idea why it is landing in Location: /usr/local/Cellar/matplotlib/2.0.0/libexec/lib/python2.7/site-packages16:23
@wikingbut ok...16:23
mikelingso I just working around this segfault for whole day and no idea why it happened and how to fix it16:24
@wikingbrew uninstall doxygen16:26
@wikingand try it w/o that16:26
mikelingok, let me have a try16:26
mikelingwiking: Oh my mine, it works.......16:36
mikelingafter remove doxygen16:36
mikelingthank you very much16:39
-!- suhas2go [uid201652@gateway/web/] has joined #shogun17:08
@sukeyNew branch feature/OutOfSource created on shogun-toolbox/shogun17:11
@sukeyNew Commit "Remove all none out of source patterns in build" to shogun-toolbox/shogun by vigsterkr:
-!- ShellcatZero [] has joined #shogun18:10
-!- HeikoS [] has joined #shogun18:15
-!- mode/#shogun [+o HeikoS] by ChanServ18:15
-!- travis-ci [] has joined #shogun18:15
travis-ciit's Viktor Gal's turn to pay the next round of drinks for the massacre he caused in shogun-toolbox/shogun:
-!- travis-ci [] has left #shogun []18:15
-!- praisethemoon [~praisethe@] has quit [Ping timeout: 248 seconds]18:33
-!- mikeling [uid89706@gateway/web/] has quit [Quit: Connection closed for inactivity]19:09
-!- HeikoS [] has quit [Ping timeout: 252 seconds]21:28
--- Log closed Tue Feb 07 00:00:14 2017