IRC logs of #shogun for Wednesday, 2016-06-22

--- Log opened Wed Jun 22 00:00:50 2016
-!- HeikoS_mobile [~Mutter@host-92-0-162-192.as43234.net] has joined #shogun		00:27
-!- HeikoS_mobile [~Mutter@host-92-0-162-192.as43234.net] has quit [Remote host closed the connection]		00:29
-!- HeikoS_mobile [~Mutter@host-92-0-162-192.as43234.net] has joined #shogun		00:56
HeikoS_mobile	bazdmeg: !!!!!!	00:56
bazdmeg	HeikoS_mobile: what	00:56
HeikoS_mobile	Nothing really ;)	00:57
HeikoS_mobile	Don't work too much	00:57
bazdmeg	haha	00:57
bazdmeg	HeikoS_mobile: my bachelor student defends something today	00:57
bazdmeg	HeikoS_mobile: spectral mixture!	00:58
-!- HeikoS_mobile [~Mutter@host-92-0-162-192.as43234.net] has quit [Read error: Connection reset by peer]		00:58
-!- HeikoS_mobile [~Mutter@host-92-0-162-192.as43234.net] has joined #shogun		00:58
bazdmeg	HeikoS_mobile: spectral mixture!	00:59
HeikoS_mobile	Kernel?	00:59
HeikoS_mobile	No please not	00:59
HeikoS_mobile	This thing is bazdmeg	00:59
HeikoS_mobile	bazdmeg^10	00:59
bazdmeg	HeikoS_mobile: 10 hours till his presentation	00:59
bazdmeg	:D	00:59
HeikoS_mobile	His?	00:59
HeikoS_mobile	Ie this	00:59
HeikoS_mobile	Or this?	00:59
bazdmeg	HeikoS_mobile: my student!	01:00
bazdmeg	bachelors	01:00
HeikoS_mobile	Spectral mixture. Man it doesn't work. You gotta tell that to someone	01:00
bazdmeg	it does!	01:00
bazdmeg	you're just lame!	01:00
bazdmeg	:D	01:00
HeikoS_mobile	Nooooo	01:00
bazdmeg	it is better than ARIMA	01:00
bazdmeg	on our data	01:00
bazdmeg	AT LEASY	01:01
HeikoS_mobile	We have a pr for it in shogun btw	01:01
bazdmeg	AT LEAST	01:01
HeikoS_mobile	Aroma?	01:01
HeikoS_mobile	This kernel is crap	01:01
HeikoS_mobile	It's a crap idea	01:01
bazdmeg	ARIMA	01:01
HeikoS_mobile	Just possible because of nips is a fucking farce	01:01
HeikoS_mobile	And all these papers are just stories that people fake	01:02
HeikoS_mobile	Spectral mixture is one of them	01:02
HeikoS_mobile	Aroma	01:02
HeikoS_mobile	What is it?	01:02
bazdmeg	autoregressive integrated moving average	01:02
HeikoS_mobile	Uh sounds fancy	01:03
HeikoS_mobile	Ha well	01:04
HeikoS_mobile	Good luck to you and your student	01:04
HeikoS_mobile	I'm sure it'll all be good	01:04
bazdmeg	HeikoS_mobile: uhmm how to show you a pic	01:05
bazdmeg	HeikoS_mobile: http://imgur.com/7BVTMG7	01:06
bazdmeg	it is not THAT bad	01:06
bazdmeg	human feature engineering like linear model on day of week	01:07
bazdmeg	is not even better	01:07
HeikoS_mobile	On toy problems like this it works fine	01:08
HeikoS_mobile	Because anything works fine there	01:08
HeikoS_mobile	Linear model on such data is bazdmeg so of course	01:08
bazdmeg	nono	01:08
bazdmeg	linear model on top of day of week	01:08
bazdmeg	would be like periodic	01:09
bazdmeg	it is quite hard to beat	01:09
HeikoS_mobile	But on any real problem, it is impossible to learn the promoters of this thing	01:09
bazdmeg	it is basically - do the same thing forever	01:09
bazdmeg	well it is real problem	01:09
bazdmeg	I mean it is real data	01:09
HeikoS_mobile	A periodic kernel plus rational quadratic as a noise model on linear regression does the same	01:09
bazdmeg	INTERNET BAZDMEG SCALE	01:09
HeikoS_mobile	But these are not interesting problems	01:09
HeikoS_mobile	Internet scale?	01:10
HeikoS_mobile	What's your N?	01:10
bazdmeg	haha well it is like visits of website	01:10
bazdmeg	I had to reduce shitload of data to get that	01:10
HeikoS_mobile	Inverting kernel matrix?	01:10
bazdmeg	haha no the data is small I am joking	01:10
bazdmeg	but you've never seen that much data	01:10
bazdmeg	I had to reduce ;)	01:10
HeikoS_mobile	We do these problems in our assignments here	01:10
HeikoS_mobile	All periodic models work well	01:11
HeikoS_mobile	Haha	01:11
HeikoS_mobile	Thing is that the marginal likelihood inherits all the periodicity	01:11
HeikoS_mobile	So you immediately get stuck when optimising it	01:11
HeikoS_mobile	Which means, you get the solution you put in initially	01:12
HeikoS_mobile	And you can't learn shit	01:12
bazdmeg	haha	01:12
bazdmeg	HeikoS_mobile: yes that's what he got	01:12
bazdmeg	but	01:12
bazdmeg	gpatt was much better	01:12
HeikoS_mobile	You know this kernel grammar by the Cambridge guys?	01:13
bazdmeg	no	01:13
HeikoS_mobile	That's much cooler	01:13
HeikoS_mobile	Lemme check	01:13
bazdmeg	kernels are shit	01:13
bazdmeg	:D	01:13
bazdmeg	show me how to run it on my scale	01:13
HeikoS_mobile	Easy	01:15
HeikoS_mobile	Btw upload your dataset to the automated statistician	01:16
HeikoS_mobile	It uses this kernel grammar	01:16
HeikoS_mobile	Gives you a latex with the same plot you showed me	01:16
HeikoS_mobile	Kinda cute	01:16
bazdmeg	yeah I know	01:16
HeikoS_mobile	And no spectral mixture kernel	01:16
bazdmeg	yeah it just	01:17
bazdmeg	like adds some sine	01:17
HeikoS_mobile	Yeah they have a generative grammar	01:17
HeikoS_mobile	And are careful about parameters	01:17
HeikoS_mobile	It also easily get linear plus periodic	01:17
HeikoS_mobile	Much more actually	01:17
HeikoS_mobile	bazdmeg: and for scale, check random features	01:18
HeikoS_mobile	Or sonney2k svm tricks	01:18
HeikoS_mobile	He did that 15 year ago	01:18
bazdmeg	HeikoS_mobile: vowpal wabbit	01:18
bazdmeg	everybody use that	01:18
HeikoS_mobile	What nips now raves about	01:18
HeikoS_mobile	We should update backend	01:18
HeikoS_mobile	I guess	01:18
HeikoS_mobile	It's good	01:18
bazdmeg	it is not	01:19
bazdmeg	but everything else is bazdmeg	01:19
HeikoS_mobile	Haha	01:19
HeikoS_mobile	You can use it with random feature kernels btw	01:19
bazdmeg	it fails badly	01:19
bazdmeg	I mean if it doesn't work	01:19
bazdmeg	it segfault or so	01:19
HeikoS_mobile	If something as easy as a linear model fails it is almost always the users fault	01:20
HeikoS_mobile	Oh	01:20
bazdmeg	but it is very versatile	01:20
HeikoS_mobile	Segfault??	01:20
HeikoS_mobile	Haha	01:20
HeikoS_mobile	Nice	01:20
bazdmeg	yeah it can segfault on bad input	01:20
HeikoS_mobile	Happy that shogun got rid of those after all	01:20
HeikoS_mobile	Had quite a few in the past	01:20
HeikoS_mobile	Btw	01:21
HeikoS_mobile	I really like your new parameter map	01:21
HeikoS_mobile	Such a relief	01:21
HeikoS_mobile	Look forward to delete Parameter.cpp	01:21
bazdmeg	well it is quite obvious	01:22
bazdmeg	it can be much more sophisticated	01:22
HeikoS_mobile	Obvious?	01:22
HeikoS_mobile	Yeah sure	01:22
HeikoS_mobile	It's the same idea as is implemented already	01:22
HeikoS_mobile	Just using c++	01:22
HeikoS_mobile	Rather than c89	01:22
bazdmeg	well 'any' is a bit fancy	01:22
bazdmeg	it is the best 'any' around there (tm)	01:22
HeikoS_mobile	You know how we did it before?	01:23
bazdmeg	yes	01:23
HeikoS_mobile	Haha :)	01:23
HeikoS_mobile	Void pointer	01:23
HeikoS_mobile	Plus enum	01:23
HeikoS_mobile	But was more powerful	01:23
HeikoS_mobile	We had strings if sgvector of sgmatrix	01:23
bazdmeg	it is a bit crazy	01:24
HeikoS_mobile	Let's delete it next week	01:24
HeikoS_mobile	After Sanuj is merged	01:25
bazdmeg	HeikoS_mobile: you can't even imagine what crazy thing I am now doing	01:25
HeikoS_mobile	Then imagine it for me	01:25
HeikoS_mobile	What isn't	01:25
HeikoS_mobile	Is it	01:25
bazdmeg	I just wrote an IPython notebook to create 18 adv campaigns	01:25
bazdmeg	:D	01:25
HeikoS_mobile	Haha	01:25
bazdmeg	it is like a plumber shit	01:26
HeikoS_mobile	Don't tell wiking he will bazdmeg you	01:26
bazdmeg	is this I was studying maths for	01:26
HeikoS_mobile	Well get a better job	01:26
bazdmeg	hahaa	01:26
bazdmeg	well	01:26
HeikoS_mobile	Don't sell yourself	01:26
bazdmeg	no I am very good there but I have to do some dirty stuff sometimes	01:26
HeikoS_mobile	To me seems like you leave a lot of life energy there	01:27
HeikoS_mobile	But can be good of course	01:27
HeikoS_mobile	Just gotta be careful	01:27
HeikoS_mobile	As getting older	01:28
bazdmeg	it is like some cost for doing something meaningful	01:28
HeikoS_mobile	And the corp doesn't give a damn about wasting their wheels, plenty new ones around	01:28
HeikoS_mobile	Well we better don't discuss this public ;)	01:28
HeikoS_mobile	Your good there	01:28
HeikoS_mobile	Long live yandex	01:29
bazdmeg	hah lets switch	01:29
HeikoS_mobile	Lol	01:29
HeikoS_mobile	Not part of my iPhone duct	01:29
HeikoS_mobile	Dict	01:29
-!- HeikoS_mobile [~Mutter@host-92-0-162-192.as43234.net] has quit [Read error: Connection reset by peer]		01:55
-!- HeikoS_mobile [~Mutter@host-92-0-162-192.as43234.net] has joined #shogun		01:56
-!- lambday [569de6b3@gateway/web/freenode/ip.86.157.230.179] has quit [Ping timeout: 250 seconds]		02:02
-!- HeikoS_mobile [~Mutter@host-92-0-162-192.as43234.net] has quit [Remote host closed the connection]		02:09
@wiking	:>>>>>>>>>>>>	02:28
bazdmeg	wiking: !	02:34
-!- besser82 [~besser82@fedora/besser82] has quit [Ping timeout: 240 seconds]		03:06
-!- GandalfTheWizard [~Eva@112.10.171.169] has joined #shogun		03:13
-!- GandalfTheWizard [~Eva@112.10.171.169] has quit [Client Quit]		03:15
-!- GandalfTheWizard [~Eva@112.10.171.169] has joined #shogun		03:25
shogun-buildbot	build #26 of memleak - valgrind is complete: Failure [failed memory check] Build details are at http://buildbot.shogun-toolbox.org/builders/memleak%20-%20valgrind/builds/26 blamelist: Heiko Strathmann <heiko.strathmann@gmail.com>, lambday <heavensdevil6909@gmail.com>, Soumyajit De <heavensdevil6909@gmail.com>	05:30
shogun-buildbot	build #1027 of nightly_none is complete: Failure [failed test] Build details are at http://buildbot.shogun-toolbox.org/builders/nightly_none/builds/1027 blamelist: Heiko Strathmann <heiko.strathmann@gmail.com>, lambday <heavensdevil6909@gmail.com>, Soumyajit De <heavensdevil6909@gmail.com>	05:36
-!- GandalfTheWizard [~Eva@112.10.171.169] has quit [Quit: Leaving.]		05:36
-!- Saurabh7 [~Saurabh7@117.202.197.85] has joined #shogun		06:14
Saurabh7	HeikoS_web, still here ? :)	06:24
Saurabh7	HeikoS_web, get_feature_vector is not working here	06:28
Saurabh7	its gives us the sample vector for an id	06:28
Saurabh7	what is needed is vector with values of a particular feature for all sample ids	06:30
shogun-buildbot	build #1157 of nightly_default is complete: Failure [failed test test_1 notebooks] Build details are at http://buildbot.shogun-toolbox.org/builders/nightly_default/builds/1157 blamelist: Heiko Strathmann <heiko.strathmann@gmail.com>, lambday <heavensdevil6909@gmail.com>, Soumyajit De <heavensdevil6909@gmail.com>	06:33
-!- OXPHOS [c68f0c0c@gateway/web/freenode/ip.198.143.12.12] has quit [Ping timeout: 250 seconds]		06:39
-!- GandalfTheWizard [~Eva@112.10.171.169] has joined #shogun		07:04
-!- OXPHOS [c68f0c0c@gateway/web/freenode/ip.198.143.12.12] has joined #shogun		08:25
-!- sanuj [~sanuj@45.127.54.49] has joined #shogun		08:42
-!- Saurabh7 [~Saurabh7@117.202.197.85] has quit [Remote host closed the connection]		08:53
-!- GandalfTheWizard [~Eva@112.10.171.169] has quit [Quit: Leaving.]		09:09
sanuj	HeikoS_web, there?	09:23
-!- besser82 [~besser82@fedora/besser82] has joined #shogun		09:43
-!- mode/#shogun [+o besser82] by ChanServ		09:43
sanuj	bazdmeg, yo	09:54
-!- HeikoS_mobile [~Mutter@82-132-216-95.dab.02.net] has joined #shogun		10:38
-!- HeikoS_mobile [~Mutter@82-132-216-95.dab.02.net] has quit [Remote host closed the connection]		10:45
-!- Saurabh7 [~saurabh@1.39.12.41] has joined #shogun		11:04
-!- HeikoS_mobile [~Mutter@82-132-216-95.dab.02.net] has joined #shogun		11:24
HeikoS_mobile	Saurabh7:	11:24
HeikoS_mobile	Soo	11:24
HeikoS_mobile	I am on phone so hard to type	11:24
HeikoS_mobile	But	11:24
HeikoS_mobile	Can't you change the loop order	11:24
Saurabh7	HeikoS_mobile, the problem is we ahve a subset	11:24
HeikoS_mobile	You do something with all the features or vector 1 then all features of vector 2	11:25
HeikoS_mobile	Etc right?	11:25
Saurabh7	we here we need all values of a particular feature	11:25
HeikoS_mobile	Sure	11:25
HeikoS_mobile	But you do that with all active vectors right?	11:25
Saurabh7	yes only with active vectors	11:26
HeikoS_mobile	And what do you do with the first feature?	11:26
HeikoS_mobile	Something that affects the others?	11:26
Saurabh7	get all values	11:26
Saurabh7	sort it	11:26
Saurabh7	start with first	11:26
Saurabh7	caluclate a criteria using it	11:26
HeikoS_mobile	Ok, then just sort all components at once	11:26
Saurabh7	compare for all features	11:26
HeikoS_mobile	This way you can go through the vectors one by one	11:27
HeikoS_mobile	Every time you touch a vector	11:27
HeikoS_mobile	Add all components to sorted set	11:27
HeikoS_mobile	One set for each feature	11:27
HeikoS_mobile	You see?	11:27
HeikoS_mobile	Rather than doing features one by one	11:27
HeikoS_mobile	You process all at once	11:27
HeikoS_mobile	And then this allows you to traverse data column wise	11:28
HeikoS_mobile	At least this is my take on it right now	11:28
HeikoS_mobile	Can you share a link with the code?	11:28
Saurabh7	but how do I sort the vlules of features	11:28
Saurabh7	this are rows	11:28
Saurabh7	while samples are columns	11:28
Saurabh7	https://github.com/shogun-toolbox/shogun/blob/develop/src/shogun/multiclass/tree/CARTree.cpp#L524	11:28
HeikoS_mobile	404	11:29
Saurabh7	what ?:D	11:29
Saurabh7	HeikoS_mobile, https://github.com/shogun-toolbox/shogun/blob/develop/src/shogun/multiclass/tree/CARTree.cpp	11:30
Saurabh7	524 line	11:30
-!- HeikoS_mobile [~Mutter@82-132-216-95.dab.02.net] has quit [Ping timeout: 246 seconds]		11:33
-!- HeikoS_mobile [~Mutter@82-132-216-95.dab.02.net] has joined #shogun		11:36
HeikoS_mobile	Back	11:36
-!- HeikoS [~heiko@82.113.183.181] has joined #shogun		11:36
-!- mode/#shogun [+o HeikoS] by ChanServ		11:36
@HeikoS	Saurabh7: jo	11:36
Saurabh7	wo	11:36
@HeikoS	now on train wifi	11:36
@HeikoS	can you send the link again?	11:36
Saurabh7	https://github.com/shogun-toolbox/shogun/blob/develop/src/shogun/multiclass/tree/CARTree.cpp#L524	11:36
@HeikoS	ok	11:37
@HeikoS	Saurabh7: so first of all	11:37
@HeikoS	feats[j]=mat(i,j);	11:38
@HeikoS	this give a double copy	11:38
@HeikoS	as mat already is a copy	11:38
Saurabh7	yes	11:38
Saurabh7	but this has to be done somewhere	11:39
Saurabh7	I mean if we want to access rows	11:39
Saurabh7	but double copy is bad	11:39
-!- HeikoS_mobile [~Mutter@82-132-216-95.dab.02.net] has quit [Remote host closed the connection]		11:39
@HeikoS	Saurabh7: so the main thing where you need all data is the sorting	11:40
Saurabh7	yes and then	11:40
Saurabh7	do some calculation	11:41
Saurabh7	but yeah the sorting is main	11:41
@HeikoS	Saurabh7: very bad :)	11:41
@HeikoS	as doesnt help at all	11:41
@HeikoS	only costs	11:41
@HeikoS	but lets see this sorting thing	11:41
@HeikoS	Saurabh7: so my point is	11:41
@HeikoS	the outer loop is over features right?	11:41
@HeikoS	and all loop iterations are independent	11:41
@HeikoS	right?	11:41
@HeikoS	could be done in parallel?	11:41
@HeikoS	this means	11:41
@HeikoS	that you can think of this differently	11:42
@HeikoS	rather than doing all features one after another	11:42
@HeikoS	process all of them at once	11:42
@HeikoS	so no loop over features outside	11:42
@HeikoS	but a loop over vectors	11:42
Saurabh7	that will be same as transpose	11:42
@HeikoS	nonon	11:42
@HeikoS	it is more a different way of thinking about the computation that is being done	11:43
@HeikoS	if all iterations of the loop are independent	11:43
@HeikoS	and inside one iteration of the loop, you touch all vectors (that particular feature of all vectors)	11:43
@HeikoS	then you can instead do	11:44
@HeikoS	you iterate over all vectors in the outer loop	11:44
@HeikoS	and inside one iteration, you "update" the things you do for all components of the vector	11:44
Saurabh7	right now the update is only getting the feature value	11:45
@HeikoS	needs quite a bit of re-factoring the code	11:45
@HeikoS	but first it is important to understand the high level idea	11:45
@HeikoS	of processing all features at once	11:45
@HeikoS	rather than one after another	11:45
Saurabh7	so if we loop over all vectors in outer loop	11:46
@HeikoS	Saurabh7: ok lets simplify a bit	11:46
@HeikoS	I want to get the idea across	11:46
@HeikoS	it is independent of RF	11:46
@HeikoS	think you have a big matrix	11:46
Saurabh7	ok	11:46
@HeikoS	you divide it into 4 blocks	11:46
@HeikoS	now you want to compute the sum of each block	11:46
@HeikoS	you get 4 numbers	11:46
@HeikoS	ok?	11:46
@HeikoS	so far so good	11:47
Saurabh7	ok :)	11:47
@HeikoS	now: you want to permute the matrix, and do the same thign again	11:47
@HeikoS	gives you 4 more numbers	11:47
-!- sanuj [~sanuj@45.127.54.49] has quit [Ping timeout: 252 seconds]		11:47
@HeikoS	now you want to do this 100s of times	11:47
Saurabh7	yep	11:47
@HeikoS	which gives you 100s of 4 numbers	11:47
@HeikoS	get it?	11:47
@HeikoS	ok	11:47
@HeikoS	rather than having a loop with 100 iterations	11:47
@HeikoS	where inside, you compute the block wise sums	11:47
@HeikoS	(which is the naive way)	11:48
@HeikoS	you can instead do	11:48
@HeikoS	you iterate through the matrix once	11:48
@HeikoS	and then for every element you touch, you figure out (for each permutation of the big matrix) to which of the 4 numbers it is added (for all 100 permutations)	11:48
@HeikoS	so then you have a single pass over the matrix	11:49
@HeikoS	and compute block wise sums for 100 permutation	11:49
@HeikoS	s	11:49
@HeikoS	get that?	11:49
Saurabh7	ok i see	11:49
@HeikoS	and it is way faster, as you can go through the matrix in memory order, rather than kind of random access from the permutations	11:50
@HeikoS	so now for the RF	11:50
@HeikoS	similar idea	11:50
@HeikoS	currently, you have a loop over all feautres, and you do something to them, every loop iteration is independent	11:50
@HeikoS	and one loop iteration touches the particular feature of all vectors	11:50
Saurabh7	yes	11:50
@HeikoS	instead, you can iterate through the vectors, and then update (just like the 100x4 terms in the sum before) the part for the corresponding features	11:51
@HeikoS	this way, you cann access data in column wise form	11:51
@HeikoS	just need to preallocate some things	11:51
@HeikoS	just lik ebefore	11:51
@HeikoS	we needed to pre-allocate the 4 numbers for each of the 100 permutations	11:51
@HeikoS	you can ask lambday if you have more questions on this	11:51
@HeikoS	it is a tricky one, but might help	11:51
@HeikoS	a lot in fact	11:51
@HeikoS	Saurabh7: back in 5 mins	11:51
-!- HeikoS [~heiko@82.113.183.181] has quit [Quit: Leaving.]		11:52
Saurabh7	yes I understand the idea	11:52
-!- HeikoS [~heiko@82.113.183.181] has joined #shogun		11:54
-!- mode/#shogun [+o HeikoS] by ChanServ		11:54
@HeikoS	Saurabh7: so the sorting is the issue	11:54
@HeikoS	Saurabh7: doing that is hard	11:54
@HeikoS	as you need to touch all	11:54
@HeikoS	but once things are sorted, should be easier	11:55
@HeikoS	Saurabh7: so I suggest	11:55
@HeikoS	in the train call	11:55
@HeikoS	you create a transposed copy of the features	11:55
@HeikoS	and then sort each column (that is sortig features)	11:55
@HeikoS	and then store that somewhere	11:56
Saurabh7	HeikoS, ok yeah thats one way with transpose	11:56
@HeikoS	but this is different to whats currently happening	11:56
@HeikoS	as you re-sort many times here	11:56
@HeikoS	I mean, you can just sort once at the start	11:56
Saurabh7	HeikoS, no there is no re sort	11:57
@HeikoS	and then you can use that sorted matrix everywhere	11:57
@HeikoS	you just have to mask it	11:57
@HeikoS	oh	11:57
@HeikoS	sorry	11:57
@HeikoS	let me read	11:57
Saurabh7	HeikoS, there is no resort here	11:57
Saurabh7	each row is accesed once	11:57
@HeikoS	you only compute the best feature once?	11:57
Saurabh7	that row is sorted and the index are traveersed	11:58
@HeikoS	I thought you did that for multiple subsets	11:58
@HeikoS	in the RF	11:58
Saurabh7	no	11:58
Saurabh7	this thing happens once	11:58
Saurabh7	on the indices left out	11:58
Saurabh7	we do it again	11:58
@HeikoS	you were talking about changing subsets before?	11:59
Saurabh7	HeikoS, https://github.com/shogun-toolbox/shogun/blob/develop/src/shogun/multiclass/tree/CARTree.cpp#L437	11:59
@HeikoS	I mean this is cart tree	11:59
@HeikoS	yeah	11:59
@HeikoS	but if you have a RF	11:59
@HeikoS	dont you do all this multiple times?	11:59
Saurabh7	ah in rf yes	11:59
@HeikoS	ok	11:59
Saurabh7	yes we will do multiple times	11:59
@HeikoS	so my point is	11:59
Saurabh7	oh yes taht would be good now isee :)	11:59
@HeikoS	precompute the sorting	11:59
@HeikoS	to-reuse in CARTtree	12:00
Saurabh7	ok now it makes sense	12:00
@HeikoS	and then inside cart-tree	12:00
@HeikoS	assume you have the transposed feature matrix sorted	12:00
@HeikoS	and then go through that vector wise	12:00
Saurabh7	yes yes thats a good idea	12:00
@HeikoS	applying the trick i mentioned	12:00
@HeikoS	so this is two steps now	12:00
@HeikoS	1.) precompute sorted transposed matrix (refactoring the existing code to use that), not changing loop order yet, passing tests	12:01
@HeikoS	2.) change loop order (f still needed)	12:01
@HeikoS	Saurabh7: sorting and subsets are somehow redundant right?	12:01
@HeikoS	if you sort all once at beginning	12:01
@HeikoS	then applying a subset doesnt change the fact that things are sorted	12:02
Saurabh7	oh but we need to sort only form active subsets	12:02
@HeikoS	as long as subset indices are also sorted	12:02
@HeikoS	yeah but you can mask that easily	12:02
@HeikoS	like imagine your data is	12:02
@HeikoS	[4,3,2,1]	12:02
@HeikoS	you sort	12:02
@HeikoS	[1,2,3,4]	12:02
Saurabh7	ye skip which are niot needed	12:02
@HeikoS	you apply subset [0,2,3] and get [1,3,4]	12:02
@HeikoS	cool	12:03
@HeikoS	and if you sort the transposed matrix	12:03
@HeikoS	you get column wise access for free	12:03
@HeikoS	BUT I think, you can still do something about your loop	12:03
@HeikoS	which call in CART/RF takes most time? did you profile?	12:03
Saurabh7	sort	12:04
Saurabh7	HeikoS, that is here https://github.com/shogun-toolbox/shogun/blob/develop/src/shogun/multiclass/tree/CARTree.cpp#L486	12:04
Saurabh7	same one	12:04
@HeikoS	ok then	12:05
-!- sanuj [~sanuj@45.127.54.49] has joined #shogun		12:05
@HeikoS	then the speedup for RF should be x-times this is called	12:05
@HeikoS	and if you iterate in memory order, your cache misses will drop a lot	12:05
@HeikoS	so thats a good start	12:05
@HeikoS	Saurabh7: happy?	12:06
Saurabh7	HeikoS, yes :)	12:06
@wiking	muy bien	12:06
Saurabh7	HeikoS, the masking is not straight forward tho	12:06
Saurabh7	for ex above we want 1,2, 4 not 1,3, 4 as output	12:07
@HeikoS	wiking: jo	12:08
@HeikoS	Saurabh7: should not be too hard	12:08
-!- sanuj [~sanuj@45.127.54.49] has quit [Client Quit]		12:08
@HeikoS	you can use a subset stack	12:08
@HeikoS	or do by hand	12:08
@HeikoS	since we dont have subsets on matrices yet	12:08
Saurabh7	HeikoS, yes need to store indices	12:08
@HeikoS	Saurabh7: yes	12:08
@HeikoS	but keep in mind	12:08
@HeikoS	you only need to sort one matrix	12:08
@HeikoS	and then you can mask that one	12:08
@HeikoS	wiking: all good?	12:09
Saurabh7	HeikoS, ok	12:09
Saurabh7	HeikoS, thanks lot let you know how it goes	12:09
@HeikoS	cool	12:09
@HeikoS	Saurabh7: looking forward to see the changes	12:10
-!- besser82 [~besser82@fedora/besser82] has quit [Ping timeout: 276 seconds]		12:10
@HeikoS	wiking: would be cool to have thes ebenchmarks online	12:11
@wiking	tell me about it :D	12:12
OXPHOS	@HeikoS: I'm almost done - though the test is failing on LinalgBackendBase->dot(), LinalgBackendEigen3->dot() stuff..I think it's some technique details I'm not familiar with. Can I pull now so you can have a look?	12:20
@HeikoS	OXPHOS: yeah please send a PR for discussion	12:21
@HeikoS	OXPHOS: awesome!!! :)	12:21
@HeikoS	OXPHOS: I am in train so my connection might drop	12:21
@HeikoS	in fact under the ocean right now	12:21
@HeikoS	between france and england :)	12:21
OXPHOS	HeikoS: don't tell me you live in france but work in england..	12:22
@HeikoS	nono :)	12:22
@HeikoS	OXPHOS: just some mini holiday	12:22
OXPHOS	HeikoS: so convenient..!	12:24
@HeikoS	OXPHOS: indeed:)	12:24
@HeikoS	OXPHOS: dont see a PR	12:25
OXPHOS	HeikoS: I didn't find the feature branch..!	12:26
@HeikoS	OXPHOS: I can make once	12:26
@HeikoS	OXPHOS: there is none?	12:26
OXPHOS	HeikoS: only the old one	12:27
@HeikoS	OXPHOS: lemme make a new one	12:27
@HeikoS	OXPHOS: did you ever merge anything to old one?	12:27
@HeikoS	OXPHOS: because if not, just use that	12:27
OXPHOS	HeikoS: thx	12:27
OXPHOS	HeikoS: yes i did..	12:27
OXPHOS	HeikoS: or maybe you can revert that one	12:28
@HeikoS	and close your old pr	12:28
@HeikoS	now in france	12:28
@HeikoS	OXPHOS: merged somethig before=	12:28
@HeikoS	?	12:28
@HeikoS	in france now	12:28
@HeikoS	OXPHOS: no need	12:28
@HeikoS	OXPHOS: send the PR against develop for now	12:29
@HeikoS	OXPHOS: i cannot create feature branch	12:29
@HeikoS	OXPHOS: I asked wiking	12:29
@HeikoS	he will create one from develop	12:29
@HeikoS	OXPHOS: he will do in 30 mins	12:29
@HeikoS	OXPHOS: for now, send to develop	12:29
@HeikoS	and I can discuss there already with you	12:30
@HeikoS	since I got a bit of time in train left	12:30
OXPHOS	HeikoS: https://github.com/shogun-toolbox/shogun/pull/3310	12:30
@HeikoS	checking	12:32
@HeikoS	OXPHOS: good pace that was :)	12:32
@wiking	mmm	12:32
@HeikoS	train internet is too slow for http ;)	12:33
@HeikoS	loaaaading....	12:33
@wiking	wwww	12:35
@HeikoS	OXPHOS: wish I could look at it ;)	12:36
@HeikoS	f5ing	12:36
@HeikoS	bazdmeg.github.com	12:36
OXPHOS	HeikoS: still loading? xd	12:37
OXPHOS	wiking: no macro?	12:37
@HeikoS	OXPHOS: it never started	12:38
@HeikoS	French countryside	12:38
@HeikoS	love irc	12:38
@HeikoS	for its lo-fi minimal bandwidth	12:38
@HeikoS	works everywhere	12:38
@wiking	:>	12:38
@wiking	here we go	12:38
@wiking	bazdmeg.github.com	12:39
@wiking	I WANT	12:39
@wiking	:DD	12:39
@wiking	how? :)	12:39
OXPHOS	it's funny i have to use vpn to log onto irc	12:39
@wiking	or bazdmeg.github.io	12:39
@wiking	OXPHOS: .cn?	12:39
OXPHOS	but i don't need it anymore once i logged in	12:39
OXPHOS	yep	12:39
@wiking	HeikoS: ping pong: "	12:39
@wiking	do we need this? i mean wasn't the idea for post GSOC stuff was that we will require c++11? @karlnapf	12:39
@wiking	"	12:40
@wiking	0/1	12:40
@wiking	?	12:40
@wiking	where is this mister i have irc everywhere?	12:40
@HeikoS	macro?	12:40
@HeikoS	cant see the file still	12:40
@HeikoS	bazdmeg!!!!!!!!!!!!!!!!	12:40
@HeikoS	maybe you can paste the PR in here ;)	12:41
@HeikoS	c++11 +1	12:41
@HeikoS	but wanted to do a release before	12:41
@HeikoS	wiking: can we do the templating more elgant?	12:41
@HeikoS	was sergeys idea	12:41
bazdmeg	whatz	12:41
@wiking	HeikoS: with linalg stuff in it?	12:42
@wiking	bazdmeg: can we has bazdmeg.github.io? :)	12:42
@wiking	https://github.com/bazdmeg	12:42
@wiking	404	12:42
@wiking	hooohohohooooooooooooooooo	12:42
@wiking	time to register :)	12:42
OXPHOS	lol	12:42
@wiking	HeikoS: so all this stuff that OXPHOS will be doing will only be part post that release no?	12:43
bazdmeg	yes lets register	12:43
OXPHOS	HeikoS: i'm now doing in init.cpp: #ifdef HAVE_CXX11 if (!sg_linalg) sg_linalg = std::unique_ptr<SGLinalg>(new shogun::SGLinalg()); #endif	12:44
@wiking	https://github.com/bazdmeg	12:44
@wiking	:)))	12:44
@wiking	so	12:44
@wiking	how do i create a github.io page bazdmeg ?:)	12:44
bazdmeg	uhmm there is some way	12:45
bazdmeg	you craete some branch	12:45
bazdmeg	create repo	12:45
bazdmeg	bazdmeg.github.io	12:45
@HeikoS	I wonder whether I get all msgs here	12:45
@HeikoS	wiking: you seem to be scattering ;)	12:45
@HeikoS	wiking: yeah hopefully, though we can already put it in now	12:45
@HeikoS	show c++11 cleaner way and I am all up for it	12:45
@HeikoS	f55555	12:45
bazdmeg	and put index.html with BAZDMEG	12:45
@HeikoS	I guess this is about doing the backend calls	12:45
@HeikoS	I dont like register, like this template more	12:45
@HeikoS	but i like c++11	12:45
@HeikoS	so share the wisdom	12:45
@HeikoS	this is bazdmeg approved	12:45
@HeikoS	OXPHOS: I have to look later	12:45
@HeikoS	cant discuss right now	12:45
OXPHOS	HeikoS: also all linalg classes are guarded by #ifdef HAVE_CXX11 because I want to stay out of the scope of the raw ptr police ;)	12:46
@HeikoS	sure	12:46
OXPHOS	HeikoS: sure! i think wiking cares mainly about the C++11 and the macros	12:46
@HeikoS	yep	12:46
@HeikoS	seems	12:46
@HeikoS	but he speaks in riddles	12:46
@HeikoS	scatwiking	12:47
@wiking	no	12:47
@wiking	wiking mainly cares about the stupid macros for templates	12:47
@wiking	:)	12:47
@wiking	that is sooooo 90s	12:47
@wiking	sooooooooooooooooooooooooo 90s	12:47
@HeikoS	I asked 100 times for draft	12:47
@HeikoS	so this is it, until there is an explcit draft	12:47
@wiking	it is even more 90s than the friend series	12:47
@wiking	i never saw there this macro hack	12:48
@HeikoS	can also copy paste it	12:48
@HeikoS	haha	12:48
@wiking	which draft had this fucking macro hack?	12:48
@wiking	you just said there's a problem	12:48
@HeikoS	wiking: so I sent it days ago	12:48
@wiking	with templated params	12:48
@wiking	and virtual functions	12:48
@wiking	that's all	12:48
@HeikoS	yesterday the final one	12:48
@HeikoS	but idea was ther earlire	12:48
@wiking	but i didn't see this macrofucking hack	12:48
@wiking	idea	12:48
@HeikoS	yeah templated virtual functions	12:48
@wiking	idea	12:48
@wiking	explicity fuck	12:48
@wiking	EXPLICIT bazdmeg	12:49
@wiking	na mind1	12:49
OXPHOS	okay let's do c++17 magic..if any!	12:49
@wiking	i'lls end the correction	12:49
@HeikoS	wiking: show	12:49
@wiking	no need c++17	12:49
@HeikoS	show show show	12:49
@wiking	11 has it	12:49
@HeikoS	cool	12:49
@HeikoS	I dont like it either	12:49
@HeikoS	so give better way	12:49
@wiking	hahaha	12:49
@wiking	sooooooooo fucking funny	12:49
@wiking	std::is gives	12:49
@wiking	About Sexually Transmitted Diseases (STDs) - KidsHealth	12:49
@wiking	anyhow lemme dig it out	12:50
bazdmeg	yes exactly	12:50
@HeikoS	lol	12:50
@HeikoS	cool	12:50
@HeikoS	should be an easy change	12:50
@HeikoS	as long as still virtual call of base pointer	12:50
@HeikoS	ah this inet	12:52
@HeikoS	OXPHOS: I gotta check it later	12:52
@HeikoS	sorry	12:52
@wiking	OXPHOS: fix your editor	12:52
@wiking	for the weird indents	12:52
OXPHOS	HeikoS: it's okay. going for dinner soon. also a basic question - sg_linalg->set_gpu_backend(new LinalgBackendViennaCL());	12:52
OXPHOS	when will this pointer be released?	12:52
@HeikoS	OXPHOS: dont know, details for now	12:53
@wiking	when you set a new one	12:53
OXPHOS	wiking: i tried. i fix with vim everytime afterwards	12:53
@wiking	OXPHOS: what r u using for editor?	12:53
OXPHOS	wiking: atom. recommended by many	12:53
@wiking	or when the SGLinalg's dtor is called	12:53
bazdmeg	OHOHH	12:53
@HeikoS	OXPHOS: there is some default backend (eigen3) and then user can change it, thats when the old one goes	12:53
@wiking	OXPHOS: hahah vim and emacs is recommended by many	12:53
bazdmeg	my first bachelor student	12:53
@wiking	atom is not many	12:53
@HeikoS	yep	12:53
@HeikoS	bazdmeg: failed? ;:D	12:53
bazdmeg	defending just now	12:53
bazdmeg	:D	12:53
OXPHOS	vim is too geek for me..	12:54
bazdmeg	in the process	12:54
@wiking	bazdmeg: :)	12:54
@wiking	OXPHOS: emacs :>	12:54
OXPHOS	gl!	12:54
@HeikoS	OXPHOS: gedit	12:54
OXPHOS	lemme try	12:54
@wiking	you just need to find an operating system for it :)	12:54
@HeikoS	WordPad	12:54
@HeikoS	notepad++	12:54
@HeikoS	hahaha	12:54
bazdmeg	total commander	12:54
@HeikoS	amigaShell	12:54
@wiking	OXPHOS: but yeah any time you set_*_backend	12:54
@wiking	you have to release the current one	12:55
@wiking	or when the dtor is called	12:55
OXPHOS	was thinking using unique_ptr for gpubackend	12:55
@wiking	that reminds meg	12:55
OXPHOS	so when set..end is called	12:55
OXPHOS	i'll copy the backend* to unique_ptr	12:55
@wiking	OXPHOS: https://github.com/shogun-toolbox/shogun/pull/3310/files#r68033111	12:56
@wiking	i mean no matter wtf	12:56
@wiking	you gotta have a dtor	12:56
@wiking	to release	12:56
@wiking	LinalgBackendBase* cpu_backend;	12:56
@wiking	LinalgBackendBase* gpu_backend;	12:56
@wiking	right?	12:56
OXPHOS	yes for now but no for unique_ptr	12:57
@wiking	i mean those pointers cannot be let there	12:57
OXPHOS	or just a default one	12:57
@wiking	?	12:57
@wiking	que?	12:57
@wiking	i mean if you have those 2 pointers	12:57
@wiking	you should release them	12:57
OXPHOS	you are right i should	12:58
@wiking	once you release SGLinalg	12:58
@wiking	period	12:58
@wiking	:)	12:58
@wiking	memory is your friend	12:58
@wiking	:>	12:58
OXPHOS	you sure?	12:58
OXPHOS	i'll take that ;)	12:58
@wiking	yeah memory is your friend	12:59
@wiking	jvm is not	12:59
@wiking	:)	12:59
@wiking	fuck the gc	12:59
@wiking	OXPHOS: so for the time being you opted for unaligned memory right?	12:59
@wiking	and you'll fix that laterz?	13:00
OXPHOS	wiking: i can fix that once this thing works	13:00
@wiking	kk	13:00
@wiking	commented some stuff	13:00
@wiking	more than some	13:00
@wiking	so fix those minors	13:00
OXPHOS	wiking: sure thx!	13:02
@HeikoS	ook I am off	13:03
@HeikoS	this doesnt lead anywhere	13:03
@HeikoS	see you later	13:04
@HeikoS	OXPHOS: saw my email?	13:04
@wiking	HeikoS: what does?	13:04
@wiking	or doesn't?	13:04
@wiking	your internet?	13:04
@HeikoS	wiking: f55555	13:04
@wiking	bazdmeg what? :D	13:04
@HeikoS	YEAP :)	13:04
@wiking	heheh ok	13:04
@wiking	viva la france	13:04
@wiking	and the fourier transform	13:04
@wiking	otherwise we wouldnt be talkign here :D	13:04
@HeikoS	bazdmeg bazdmegstolovic	13:05
@HeikoS	yes fourier	13:05
@HeikoS	lalala	13:05
@HeikoS	bye :)	13:05
@HeikoS	ill get back later with wine and cheese	13:05
OXPHOS	HeikoS: only the midterm one	13:05
@wiking	OXPHOS: plz use internal printers, not std::cerr	13:05
@HeikoS	yeah	13:05
@HeikoS	OXPHOS: thats the one I meant	13:05
@HeikoS	+1	13:05
OXPHOS	SG_SWARNING didn't print anything for me	13:05
OXPHOS	any idea why?	13:05
@wiking	gdb/lldb is your other friend	13:06
@HeikoS	OXPHOS: you gotta activate	13:06
@HeikoS	sg_obj->get_gloval_io()->set_loglevel(MSG_WARNING)	13:06
@wiking	do you need help how to use it?	13:06
OXPHOS	will be appreciate!	13:06
@wiking	ok so	13:07
@wiking	gdb ./program	13:07
@wiking	b Class::function	13:07
@wiking	or	13:07
@wiking	b Filename:<line number>	13:07
@HeikoS	or MSG_DEBUG	13:07
@HeikoS	ah no	13:07
@HeikoS	warning should always be printed	13:07
@HeikoS	anyways	13:07
@HeikoS	bye	13:07
@wiking	and then	13:08
@wiking	r	13:08
OXPHOS	bye	13:08
@wiking	and then you can use	13:08
@wiking	s = as step	13:08
@wiking	and n = as next	13:08
@wiking	to walk in the execution	13:08
OXPHOS	does [gdb ./program] suppose to run in terminal?	13:08
@wiking	yesp	13:08
OXPHOS	command not found. guess i'll install it first	13:09
@wiking	apt-get instlal gdb	13:09
@wiking	btw apt-get moo is your friend	13:09
OXPHOS	also, i tried to brew ccache today, but failed..can't find some ascii doc	13:09
-!- HeikoS [~heiko@82.113.183.181] has quit [Remote host closed the connection]		13:09
@wiking	http://linux.bytesex.org/gdb.html	13:09
OXPHOS	aah it's working!	13:11
-!- Saurabh7 [~saurabh@1.39.12.41] has quit [Quit: Leaving]		13:22
@wiking	aah bazdmeg	14:37
@wiking	OXPHOS: u r now in china right?	14:37
OXPHOS	wiking: yep	14:37
bazdmeg	bazdmeg	14:42
-!- Saurabh7 [~Saurabh7@117.202.197.85] has joined #shogun		14:48
@wiking	bazdmeg	14:51
@wiking	!	14:51
bazdmeg	wiking: 5	14:51
bazdmeg	4	14:51
bazdmeg	3	14:51
@wiking	bazdmeg: what do we want on the page? :))	14:51
bazdmeg	2	14:51
bazdmeg	1	14:51
bazdmeg	bazdmeg	14:51
@wiking	baaazdmeg!!!	14:51
bazdmeg	wiking: BAZDMEG in 72 pt	14:51
@wiking	ok	14:51
@wiking	doing it!	14:51
bazdmeg	wiking: green on blue background	14:51
bazdmeg	comic sans	14:51
@wiking	gimme css	14:53
-!- Saurabh7 [~Saurabh7@117.202.197.85] has quit [Ping timeout: 276 seconds]		15:00
@wiking	hahahaha	15:01
@wiking	bazdmeg: you sure you wanna have it like this	15:01
-!- OXPHOS [c68f0c0c@gateway/web/freenode/ip.198.143.12.12] has quit [Quit: Page closed]		15:02
bazdmeg	yes	15:02
bazdmeg	bazdmeg: where is it	15:02
@wiking	ok pushed	15:06
@wiking	https://github.com/bazdmeg/bazdmeg.github.io/blob/master/LICENSE	15:06
@wiking	dunno how long it takes to generate the site	15:06
bazdmeg	comic sans	15:06
bazdmeg	!!!!	15:06
@wiking	i hope u like the license	15:06
bazdmeg	yes	15:06
bazdmeg	wiking: 180pt! olive!	15:07
@wiking	font-family: sans-serif;	15:07
@wiking	good?	15:07
bazdmeg	comic sans	15:08
bazdmeg	font-family: "Comic Sans", sans-serif;	15:08
@wiking	mmm not good in my ff	15:09
@wiking	"Comic Sans MS", cursive, sans-serif	15:09
@wiking	no?	15:09
@wiking	pushed	15:10
bazdmeg	ahh	15:10
bazdmeg	yes	15:10
@wiking	https://bazdmeg.github.io	15:10
@wiking	\o/	15:10
@wiking	wooooohoooo	15:10
bazdmeg	uh	15:10
bazdmeg	that's good as it can be	15:10
@wiking	heheheh	15:11
bazdmeg	wiking: I am the art director of this website	15:11
@wiking	for further improvment people can send issues and prs	15:20
@wiking	;)	15:20
@wiking	:D	15:20
@wiking	bazdmeg: http://giphy.com/gifs/xT0GqpswuzhOqHP6gM	15:23
@wiking	we could post this on xmas day	15:23
@wiking	;)	15:23
-!- GandalfTheWizard [~Eva@112.10.171.169] has joined #shogun		16:02
-!- GandalfTheWizard [~Eva@112.10.171.169] has quit [Client Quit]		16:04
-!- GandalfTheWizard [~Eva@112.10.171.169] has joined #shogun		16:05
-!- sanuj [~sanuj@45.127.54.49] has joined #shogun		16:11
sanuj	bazdmeg, yo	16:11
sanuj	bazdmeg, got time?	16:12
-!- sanuj [~sanuj@45.127.54.49] has quit [Ping timeout: 276 seconds]		16:37
-!- sanuj [~sanuj@45.127.54.49] has joined #shogun		17:05
sanuj	bazdmeg, yoyo	17:06
sanuj	HeikoS_web, can you review a PR?	17:22
-!- GandalfTheWizard [~Eva@112.10.171.169] has quit [Quit: Leaving.]		18:39
CaBa_	are there still any string kernel implementations in shogun that incorporate amino acid properties?	19:00
CaBa_	sonney2k ping	19:00
sanuj	bazdmeg, ping pong	19:09
-!- sanuj [~sanuj@45.127.54.49] has quit [Quit: Leaving]		20:34
-!- c4goldsw [5da420e6@gateway/web/cgi-irc/kiwiirc.com/ip.93.164.32.230] has joined #shogun		21:26
-!- c4goldsw [5da420e6@gateway/web/cgi-irc/kiwiirc.com/ip.93.164.32.230] has quit [Client Quit]		21:30
-!- c4goldsw [5da420e6@gateway/web/cgi-irc/kiwiirc.com/ip.93.164.32.230] has joined #shogun		21:48
c4goldsw	Hello	21:48
-!- c4goldsw [5da420e6@gateway/web/cgi-irc/kiwiirc.com/ip.93.164.32.230] has quit [Quit: http://www.kiwiirc.com/ - A hand crafted IRC client]		23:10
--- Log closed Thu Jun 23 00:00:51 2016

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