--- Log opened Mon Jul 23 00:00:17 2012 | ||
blackburn | heiko: btw I've added some multitask to tutorial | 00:05 |
---|---|---|
-!- heiko [~heiko@host86-181-80-49.range86-181.btcentralplus.com] has quit [Ping timeout: 264 seconds] | 00:07 | |
gsomix | sonney2k, all news will be tomorrow. | 00:15 |
gsomix | I need to sleep a little | 00:15 |
gsomix | good night guys | 00:15 |
-!- heiko [~heiko@host86-181-80-49.range86-181.btcentralplus.com] has joined #shogun | 00:19 | |
-!- heiko [~heiko@host86-181-80-49.range86-181.btcentralplus.com] has left #shogun [] | 00:24 | |
wiking | done | 00:45 |
-!- romi_ [~mizobe@189-18-41-111.dsl.telesp.net.br] has joined #shogun | 02:10 | |
-!- puffin444 [42676f92@gateway/web/freenode/ip.66.103.111.146] has joined #shogun | 02:56 | |
blackburn | puffin444: good evening ;) | 03:08 |
puffin444 | Hey blackburn, how is it going? | 03:08 |
blackburn | pretty fine and you? | 03:08 |
puffin444 | Doing well. Almost done coding the FITC algorithm. Eigen is a lifesaver. | 03:09 |
blackburn | :D | 03:09 |
droopy | not bad | 03:09 |
blackburn | what is FITC? | 03:09 |
puffin444 | Sparse Approximation for GPs | 03:10 |
blackburn | just don't forget to guard it with HAVE_EIGEN3 | 03:10 |
puffin444 | Oh yes, will do. | 03:10 |
wiking | yeey | 03:14 |
wiking | mocked it ;) | 03:15 |
blackburn | mockingwiking | 03:16 |
wiking | blackburn: for sure ;) | 03:16 |
wiking | but now i should start pooring unit testing files into that PR | 03:17 |
wiking | so that actually it'll make it worth to merge it | 03:17 |
blackburn | it looks like you liked this idea pretty much :) | 03:17 |
wiking | yeah | 03:18 |
wiking | dunno why | 03:18 |
wiking | got into it | 03:18 |
CIA-18 | shogun: Sergey Lisitsyn master * r28824e9 / (7 files in 3 dirs): Added fused lasso capabilities to the SLEP solver - http://git.io/XZ1iLg | 04:17 |
-!- blackburn [~blackburn@109.226.92.17] has quit [Quit: Leaving.] | 04:56 | |
-!- romi_ [~mizobe@189-18-41-111.dsl.telesp.net.br] has quit [Quit: Leaving] | 06:25 | |
-!- puffin444 [42676f92@gateway/web/freenode/ip.66.103.111.146] has quit [Ping timeout: 245 seconds] | 06:55 | |
-!- uricamic [~uricamic@2001:718:2:1634:a553:a843:34d8:fea5] has joined #shogun | 08:51 | |
-!- gsomix_ [~gsomix@95.67.172.151] has joined #shogun | 09:30 | |
-!- vojtech [9320543b@gateway/web/freenode/ip.147.32.84.59] has joined #shogun | 09:32 | |
-!- gsomix [~gsomix@88.200.214.225] has quit [Ping timeout: 248 seconds] | 09:33 | |
-!- n4nd0 [~nando@s83-179-44-135.cust.tele2.se] has joined #shogun | 09:36 | |
-!- pluskid [~pluskid@111.120.44.221] has joined #shogun | 10:02 | |
gsomix_ | good morning | 10:11 |
droopy | roger!?! | 10:11 |
-!- gsomix_ [~gsomix@95.67.172.151] has quit [Ping timeout: 240 seconds] | 10:29 | |
-!- gsomix_ [~gsomix@80.234.56.113] has joined #shogun | 10:41 | |
-!- pluskid [~pluskid@111.120.44.221] has quit [Ping timeout: 248 seconds] | 10:52 | |
-!- pluskid [~pluskid@202.130.113.141] has joined #shogun | 10:52 | |
-!- alexlovesdata [~binder@goldenezahl.ml.tu-berlin.de] has joined #shogun | 11:14 | |
-!- n4nd0 [~nando@s83-179-44-135.cust.tele2.se] has quit [Ping timeout: 240 seconds] | 11:57 | |
-!- gsomix_ is now known as gsomix | 12:13 | |
-!- blackburn [~blackburn@109.226.92.17] has joined #shogun | 12:19 | |
-!- gsomix_ [~gsomix@109.169.141.131] has joined #shogun | 12:23 | |
-!- gsomix [~gsomix@80.234.56.113] has quit [Ping timeout: 264 seconds] | 12:25 | |
-!- n4nd0 [~nando@s83-179-44-135.cust.tele2.se] has joined #shogun | 13:05 | |
-!- blackburn [~blackburn@109.226.92.17] has quit [Quit: Leaving.] | 13:24 | |
-!- needsch [~user@ip-176-198-229-75.unitymediagroup.de] has joined #shogun | 13:45 | |
-!- pluskid_ [~pluskid@li411-226.members.linode.com] has joined #shogun | 14:04 | |
-!- pluskid [~pluskid@202.130.113.141] has quit [Ping timeout: 264 seconds] | 14:08 | |
-!- Netsplit *.net <-> *.split quits: gsomix_, n4nd0, fabs | 14:09 | |
-!- Netsplit over, joins: n4nd0, gsomix_, fabs | 14:10 | |
-!- Netsplit *.net <-> *.split quits: gsomix_, n4nd0, fabs | 14:21 | |
-!- Netsplit over, joins: n4nd0, gsomix_, fabs | 14:22 | |
wiking | n4nd0: are u a split currently? | 15:10 |
n4nd0 | wiking: a split? | 15:12 |
n4nd0 | do you mean about this netsplit ^ | 15:12 |
n4nd0 | if so, I have no idea | 15:12 |
n4nd0 | gtg now, bye! | 15:16 |
-!- n4nd0 [~nando@s83-179-44-135.cust.tele2.se] has quit [Quit: leaving] | 15:16 | |
wiking | hahah | 15:16 |
wiking | well there goes n4nd0 :) | 15:17 |
-!- pluskid_ [~pluskid@li411-226.members.linode.com] has quit [Ping timeout: 260 seconds] | 15:21 | |
-!- gsomix_ is now known as gsomix | 15:22 | |
-!- pluskid_ [~pluskid@111.120.44.221] has joined #shogun | 15:34 | |
-!- pluskid_ [~pluskid@111.120.44.221] has quit [Quit: Leaving] | 16:01 | |
-!- n4nd0 [~nando@s83-179-44-135.cust.tele2.se] has joined #shogun | 16:29 | |
-!- heiko1 [~heiko@host86-181-80-49.range86-181.btcentralplus.com] has joined #shogun | 16:35 | |
-!- gsomix [~gsomix@109.169.141.131] has quit [Ping timeout: 240 seconds] | 16:37 | |
-!- heiko2 [~heiko@host86-177-114-238.range86-177.btcentralplus.com] has joined #shogun | 16:56 | |
-!- heiko1 [~heiko@host86-181-80-49.range86-181.btcentralplus.com] has quit [Ping timeout: 264 seconds] | 16:57 | |
CIA-18 | shogun: Heiko Strathmann master * r62c71cf / (8 files in 4 dirs): Merge pull request #669 from karlnapf/master - http://git.io/JUQgpw | 17:11 |
CIA-18 | shogun: Heiko Strathmann master * r1cbdee8 / (3 files): make use of new data generator class - http://git.io/dsTzmw | 17:11 |
CIA-18 | shogun: Heiko Strathmann master * r3f6906d / (4 files in 2 dirs): made all python examples use data generator class - http://git.io/R_-Frw | 17:11 |
CIA-18 | shogun: Heiko Strathmann master * r5c16e8c / src/shogun/features/DataGenerator.h : typo - http://git.io/Z2HLaQ | 17:11 |
-!- zxtx [~zv@ool-457e7550.dyn.optonline.net] has quit [Ping timeout: 260 seconds] | 17:19 | |
-!- gsomix [~gsomix@95.67.188.241] has joined #shogun | 17:27 | |
gsomix | sonney2k, hey, thunderstorm, hurricane, again. =___= | 17:28 |
-!- vojtech [9320543b@gateway/web/freenode/ip.147.32.84.59] has quit [Ping timeout: 245 seconds] | 17:35 | |
-!- blackburn [~blackburn@109.226.92.17] has joined #shogun | 17:40 | |
-!- emrecelikten [~emre@212.253.72.103] has joined #shogun | 17:44 | |
-!- n4nd0 [~nando@s83-179-44-135.cust.tele2.se] has quit [Ping timeout: 245 seconds] | 17:54 | |
-!- ckwidmer [8ca3fe9d@gateway/web/freenode/ip.140.163.254.157] has joined #shogun | 17:57 | |
-!- heiko2 [~heiko@host86-177-114-238.range86-177.btcentralplus.com] has quit [Quit: Leaving.] | 18:34 | |
-!- uricamic [~uricamic@2001:718:2:1634:a553:a843:34d8:fea5] has quit [Ping timeout: 245 seconds] | 18:35 | |
-!- alexlovesdata [~binder@goldenezahl.ml.tu-berlin.de] has quit [Quit: Leaving.] | 19:06 | |
rieck | make your own decision: http://www.geekwire.com/2012/raunchy-windows-azure-dance-routine/ | 19:32 |
blackburn | rieck: m$ got totally crazy? :) | 19:34 |
emrecelikten | I have seen this before, but this line still gets me: "In a strange effort to be inclusive, a monitor displaying the lyrics added, "or vagina."" | 19:48 |
blackburn | I am sorry but do they mean it is good to have big and hard vagina? :D | 19:53 |
blackburn | I lose some kind of logic here | 19:55 |
emrecelikten | I think someone as stupid to organize this event might as well think anything | 19:55 |
blackburn | yeah that should be an explanation | 19:56 |
emrecelikten | http://www.kurzweilai.net/poison-attacks-against-machine-learning | 20:23 |
audy | emrecelikten i've seen this happen | 20:47 |
audy | bots posting in forums to confuse spam detectors | 20:47 |
rieck | yep, that's an interesting paper of blaine and pavel | 20:58 |
rieck | unfortunately, they do not present a defense in this work. | 20:58 |
rieck | http://arxiv.org/abs/1206.6389v1 | 21:00 |
-!- heiko1 [~heiko@host86-177-114-238.range86-177.btcentralplus.com] has joined #shogun | 21:09 | |
@sonney2k | rieck, what is the essence of this paper? | 21:12 |
@sonney2k | I am rather surprised by this finding... | 21:12 |
@sonney2k | gsomix, ohh good luck! | 21:12 |
rieck | afaik, they idea is to maximally change the hyperplane with a few points | 21:13 |
rieck | this can be cast into an optimization problem, though np-hard. | 21:13 |
rieck | however, if you relax it somewhat, you can still create havoc | 21:13 |
rieck | actually, i didn't read the paper and just talked with blaine nelson | 21:13 |
gsomix | sonney2k, hey :) it's all right | 21:13 |
rieck | i think it's not so surprising | 21:14 |
rieck | similar things have been shown by marius for simple hypersphrere models | 21:14 |
rieck | a key point with all poisoning attacks is the ratio of malicious to benign training points | 21:15 |
@sonney2k | rieck, blackburn - wtf "The words MICRO and SOFT don’t apply to my penis or vagina.”? | 21:15 |
blackburn | sonney2k: m$ that is | 21:15 |
rieck | sonney2k: i am afraid, i don't know about your penis. | 21:15 |
blackburn | or vagina | 21:15 |
@sonney2k | rieck, I would expect one needs 0.1 * nr_train_points to do poisoning | 21:15 |
emrecelikten | haha | 21:15 |
@sonney2k | rieck, sure you do :D | 21:16 |
rieck | sonney2k: marius showed for hypersphere that the amount of movement from the original decision function is linear in the number of malicious points | 21:16 |
blackburn | oh that's way too gay boys | 21:16 |
rieck | i am not sure, but i think there is a similar relation in the new paper | 21:16 |
@sonney2k | you were my office mate! | 21:16 |
rieck | that is, you can move the hyperplane by d percent if you control d percent of the data | 21:16 |
@sonney2k | yeah but that really is not disturbing then | 21:17 |
rieck | blackburn: you never wondered that only "boys" work on shogun? | 21:17 |
@sonney2k | rieck, blackburn is gay too - look at the pictures he sends around :D | 21:18 |
blackburn | rieck: well it can be explained in other terms :) | 21:18 |
rieck | sonney2k: yep. i think it came to icml, because they modeled it nicely | 21:18 |
blackburn | namely women are bad in shoguning | 21:19 |
@sonney2k | wiking, my internet connection broke down yesterday and is still flaky (here on that island) | 21:19 |
@sonney2k | wiking, so no idea what you wanted to ask / waht the status is | 21:19 |
rieck | blackburn: "shoguning"... i can imagine what that is ;) | 21:20 |
rieck | sonney2k: btw, is there a tom-tom app for android? | 21:20 |
blackburn | rieck: we also have word vodking here | 21:20 |
@sonney2k | rieck, I just now see that pavel is a co-author on that paper... | 21:21 |
blackburn | his name sounds russian | 21:22 |
rieck | he's from moscow | 21:22 |
rieck | anyway, i have to leave for dinner ... somebody is already calling my name | 21:22 |
wiking | sonney2k: oh yeah still here? | 21:23 |
@sonney2k | guys our 9th slot that we had to give away due to lack of mentors was nicely invested: http://jordi.inversethought.com/blog/octconf-2012-report/ and http://hg.savannah.gnu.org/hgweb/octave/graph/0ef151f9fdc9 | 21:23 |
@sonney2k | blackburn, ^ | 21:23 |
@sonney2k | basically octave has a JIT now :D | 21:23 |
wiking | sonney2k: so weekly is about to be written & sent | 21:23 |
wiking | sonney2k: and there was some thing about the configure script and Makefile in tests/unit but i've solved it in a way, just check the corresponding commit in the PR | 21:24 |
blackburn | yeah much more useful than our crappy stuff | 21:24 |
blackburn | :) | 21:24 |
@sonney2k | rieck, WIP AFAIK | 21:24 |
wiking | sonney2k: whether u like the solution or not | 21:24 |
wiking | sonney2k: i'll just do the changes that heiko1 told in the comments | 21:25 |
@sonney2k | blackburn, what crappy stuff? | 21:25 |
wiking | sonney2k: and let me know if there's anything else u wanna change in there | 21:25 |
blackburn | sonney2k: the one I do for example :D | 21:25 |
@sonney2k | blackburn, should we make shogun tests a git submodule? | 21:25 |
blackburn | no, why? | 21:25 |
@sonney2k | then we can give wiking r/w access on that one too | 21:26 |
emrecelikten | Look what I have done to this channel with a few pieces of text and a couple of links, everybody started talking | 21:26 |
blackburn | emrecelikten: keep going on this way :) | 21:26 |
emrecelikten | :D | 21:26 |
@sonney2k | emrecelikten, more please | 21:26 |
blackburn | sonney2k: in that case yes it makes sense | 21:26 |
@sonney2k | I mean it is wiking / heiko pushing that stuff | 21:27 |
blackburn | however is it needed? | 21:27 |
@sonney2k | so they should just do it | 21:27 |
heiko1 | sonney2k, is it hard for the build bot to compile latex? | 21:27 |
@sonney2k | heiko1, ??? | 21:27 |
@sonney2k | hard as in penis / vagina? | 21:27 |
wiking | sonney2k: heheh or u can give me r/w for shogun straight ;) | 21:27 |
heiko1 | lol ;) not exactly | 21:27 |
heiko1 | what do you guys have with that video? ;) | 21:28 |
heiko1 | sonney2k, in order to have a live version of shogun-tutorial like the latest doxygen online | 21:28 |
@sonney2k | heiko1, I just can't stop lauging about it | 21:28 |
@sonney2k | laughing | 21:28 |
heiko1 | hehe ;) | 21:28 |
droopy | 8) | 21:28 |
@sonney2k | wonder when there will be some azure memes involving P* and V*'s | 21:28 |
@sonney2k | wiking, that takes about a year :D | 21:29 |
@sonney2k | ask heiko! | 21:29 |
blackburn | sonney2k: I think it could be a nice adv | 21:29 |
wiking | heiko1: the reason i have that static const in the SGVector_unittest to avoid code repetition | 21:29 |
blackburn | our cloud is hard just like your penis | 21:29 |
blackburn | and flexible like | 21:29 |
blackburn | you know | 21:29 |
@sonney2k | and soft | 21:29 |
@sonney2k | like you know | 21:29 |
@sonney2k | runs the soft you have | 21:30 |
wiking | heiko1: imo it's just waste of lines to add everywhere that constant vector. but you are right that it should be rather randomly generated vectors | 21:30 |
heiko1 | wiking, I know, but its hard to overview when one starts doing this and since your tests will be serving as an example for others, Id like to avoid that. I had to search for the a vector when I first read it, so a lot of scrolling. And one line code-repetitions dont really hard --especially when we are in the context of testing , where coe repetition happens perse | 21:31 |
wiking | heiko1: (nod) | 21:31 |
blackburn | what the f you are talking about :D | 21:31 |
wiking | heiko1: i rather then just remove that vector and simply use randomly generated vectors. | 21:31 |
@sonney2k | isn't traveling salesmen in P* ? | 21:32 |
heiko1 | sonney2k, lol :D | 21:32 |
heiko1 | its in VP ;) | 21:32 |
@sonney2k | ohh no I think it is hard | 21:32 |
@sonney2k | NP hard | 21:32 |
wiking | ;> | 21:32 |
blackburn | :D | 21:32 |
droopy | ;) | 21:32 |
wiking | i think it's even complete... | 21:33 |
heiko1 | wiking, great work btw on the testing stuff, I am really looking forward to that! | 21:33 |
@sonney2k | heiko1, so what did you want to ask | 21:33 |
emrecelikten | Laughing out loud here | 21:33 |
wiking | heiko1: yeah i think we shoudl just roll out very fast a lot of tests | 21:33 |
@sonney2k | blackburn, ok I am creating a submodule then | 21:33 |
emrecelikten | NP hard | 21:33 |
heiko1 | sonney2k, would be nice to have a live version of shogun-tutorial like the latest doxygen online | 21:34 |
@sonney2k | heiko1, that is no problem | 21:34 |
blackburn | btw I expect it grow up to 100-150 pages pretty fast | 21:34 |
@sonney2k | only my network connection might make this impossible so in worst case it will have to wait 1 week | 21:35 |
blackburn | in case of such big blank fields like heiko1 wants :D | 21:35 |
heiko1 | sonney2k, then go for it! would be cool! | 21:35 |
heiko1 | sonney2k ok | 21:35 |
@sonney2k | heiko1, latex is already on the build server - doxygen uses latex to generate the formulas | 21:35 |
heiko1 | blackburn, but I like your font :) | 21:35 |
heiko1 | sonney2k, cool, then go! | 21:35 |
@sonney2k | blackburn, did you already make a submodule out of the tutorial? | 21:35 |
blackburn | nope | 21:35 |
blackburn | I'll let you do the job | 21:36 |
blackburn | :D | 21:36 |
@sonney2k | blackburn, it will be shogun/tutorial right? | 21:36 |
blackburn | exactly | 21:36 |
@sonney2k | or shogun/doc/tutorial? | 21:36 |
@sonney2k | what? | 21:36 |
@sonney2k | or shogun/manual? | 21:37 |
blackburn | shogun/tutorial | 21:37 |
@sonney2k | ok | 21:37 |
blackburn | manual is for soft P* | 21:37 |
@sonney2k | blackburn, shoudl we move the other stuff then? | 21:37 |
blackburn | which stuff? | 21:37 |
@sonney2k | I mean to shogun/doc/reference | 21:37 |
@sonney2k | the doxygen doc | 21:37 |
blackburn | hmm | 21:37 |
@sonney2k | I mean it is a reference documentation | 21:38 |
blackburn | hmm | 21:38 |
blackburn | no idea | 21:38 |
@sonney2k | by no means a tutorial | 21:38 |
@sonney2k | bah | 21:38 |
blackburn | I am ok with current structure | 21:38 |
blackburn | just add shogun/tutorial | 21:38 |
heiko1 | yeah just add | 21:39 |
@sonney2k | but it is cluttered | 21:39 |
@sonney2k | we have shogun/doc/images | 21:39 |
@sonney2k | we have shogun/doc/pages{,_cn} | 21:39 |
@sonney2k | and some scripts in there | 21:39 |
blackburn | tutorial has some cluttered structure as well | 21:40 |
blackburn | so? | 21:40 |
blackburn | :) | 21:40 |
@sonney2k | I will create shogun/doc/tutorial | 21:41 |
@sonney2k | I will create shogun/doc/reference | 21:41 |
blackburn | I will kill all humans | 21:43 |
blackburn | all your bayes are belong to us | 21:43 |
blackburn | sonney2k: are you still in mood to write up something? | 21:47 |
blackburn | we started already | 21:48 |
wiking | ok so heiko1 | 21:48 |
wiking | we have a problem :))) | 21:48 |
wiking | lib/SGVector_unittest.cc:65: Failure | 21:48 |
wiking | Value of: a.dot(a.vector,a.vector, a.vlen) Actual: 3521657.9835528652 | 21:48 |
wiking | Expected: dot_val | 21:48 |
wiking | Which is: 3521657.9835528648 | 21:48 |
heiko1 | ha | 21:48 |
heiko1 | wiking, random values? :) | 21:48 |
wiking | yes | 21:48 |
heiko1 | why does that happen? | 21:48 |
heiko1 | where do types change? | 21:48 |
wiking | well rounding error somewher | 21:48 |
wiking | heiko1: nowhere | 21:48 |
heiko1 | mmh, ok | 21:49 |
wiking | heiko1: so maybe we can say | 21:49 |
heiko1 | but the multiplication has to happen in exactly the same way | 21:49 |
heiko1 | if you do things like a+a | 21:49 |
wiking | fabs(expected-got, 10E-6) | 21:49 |
blackburn | iirc double is for 8 valid digits | 21:49 |
heiko1 | and then check for 2*a | 21:49 |
wiking | i mean | 21:49 |
blackburn | am I wrong? | 21:49 |
wiking | fabs(expected-got) < 10e-6 then we should accept | 21:49 |
heiko1 | wiking, 10E-6 is not enough | 21:50 |
blackburn | heiko1: are you constructing moonwalkers? | 21:50 |
blackburn | :D | 21:50 |
heiko1 | Immh, I didnt really think of this :) | 21:50 |
@sonney2k | blackburn, no it is true | 21:50 |
blackburn | sonney2k: about digits? | 21:50 |
@sonney2k | double should be 1e-12 to 1e-16 | 21:50 |
heiko1 | I kind of had integer number in mind when I suggested this | 21:50 |
wiking | heiko1: well if SGVector<float64_t> | 21:51 |
heiko1 | wiking, yeah, sorry my fault | 21:51 |
wiking | then i guess it's completely valid to say that it'll be any stupid float64_t | 21:51 |
heiko1 | yeah | 21:51 |
wiking | so we can go really with some error | 21:51 |
wiking | and that's it | 21:51 |
heiko1 | do it as precise as you can and then I think its fine | 21:51 |
wiking | mmmhehehe | 21:51 |
blackburn | wiking: other thing you may do is to reduce vectors | 21:51 |
heiko1 | since errors in implementation usually do not pop up in 10E-6 domains :) | 21:51 |
blackburn | 10E6 rather | 21:52 |
heiko1 | but make sure that it always passes. I had the problem with some of my tests that they passed only 99% of the time | 21:52 |
blackburn | hah | 21:52 |
blackburn | some DR algorithm still passes in 99% of all time | 21:52 |
blackburn | kernel LTSA IIRC | 21:53 |
wiking | ok this one goes ok: | 21:53 |
wiking | float64_t error = CMath::abs (dot_val - a.dot(a.vector,a.vector, a.vlen)); 66 EXPECT_TRUE(error < 10E-10); | 21:53 |
blackburn | heiko1: I have one issue with my multishit | 21:54 |
heiko1 | wiking, ok! | 21:54 |
heiko1 | blackburn, whats the problem | 21:54 |
heiko1 | blackburn, are you ok with merging the test framework when wiking finished the latest things? | 21:54 |
blackburn | heiko1: sure | 21:54 |
blackburn | heiko1: ok the problem is | 21:54 |
blackburn | to declare what tasks you are working on | 21:54 |
blackburn | I provide Task class | 21:54 |
blackburn | which is basically indices vector | 21:55 |
blackburn | currently two task relations are supported | 21:55 |
blackburn | TaskGroup and TaskTree | 21:55 |
blackburn | just consider it as set of Tasks(indices vectors) | 21:55 |
blackburn | so the problem is how to handle all that sh with subsets | 21:56 |
heiko1 | k | 21:56 |
heiko1 | phew | 21:56 |
heiko1 | complicated | 21:56 |
gsomix | sonney2k, ta dam! | 21:56 |
blackburn | yeah it is | 21:56 |
@sonney2k | wiking, hmmhh 1e-10 ? won't 1e-12 to 1e-15 work too? | 21:56 |
@sonney2k | gsomix, ohh I like that sound :D | 21:57 |
heiko1 | blackburn, I still think that the only solution that would fit into current framework would be to create a new splitting strategy class which can handle all these index sets in the right way | 21:57 |
wiking | sonney2k: did work... | 21:57 |
wiking | tell me a number to put there and i'll put it there | 21:58 |
blackburn | heiko1: no, it is even more complex | 22:00 |
blackburn | heiko1: I think one solution | 22:01 |
blackburn | could be | 22:01 |
blackburn | to obtain real indices when training | 22:01 |
blackburn | i.e. to put only vectors that are in | 22:01 |
heiko1 | blackburn, so you need to know the result of the training before constructing indices? | 22:02 |
heiko1 | blackburn, afk 5 mins | 22:02 |
blackburn | no, why | 22:02 |
blackburn | example | 22:02 |
blackburn | tasks are | 22:02 |
blackburn | [0,1,2] [3,4,5] | 22:03 |
blackburn | and subset is | 22:03 |
blackburn | [0,1,4] | 22:03 |
blackburn | so what can I come up with | 22:03 |
blackburn | is to create vectors | 22:03 |
blackburn | [0,1] [4] | 22:03 |
blackburn | to describe these two tasks | 22:03 |
gsomix | sonney2k, one moment. need update pr | 22:07 |
@sonney2k | wiking, I think 1e-15 should be fine | 22:08 |
@sonney2k | gsomix, k | 22:08 |
wiking | sonney2k: fails... | 22:08 |
blackburn | sonney2k: doesn't it depend on scale? | 22:08 |
wiking | sonney2k: now it failed with the given random vector and using 10e-12 | 22:08 |
@sonney2k | wiking, how big are the numbers? | 22:20 |
@sonney2k | wiking, can you check for *relative* precision? | 22:20 |
@sonney2k | i.e. |a-b|/(a+b) < 1e-12 ... | 22:20 |
wiking | ok i dont get this... | 22:21 |
wiking | sonney2k: jsut a sec | 22:21 |
wiking | SGVector<float64_t>::scale_vector(-1.0, a.vector, a.vlen); | 22:21 |
wiking | EXPECT_EQ(sum, SGVector<float64_t>::sum_abs(a.vector, a.vlen)); | 22:21 |
heiko1 | blackburn, re, sorry took longer since the shop-owner wanted to break his Ramadan with me :) | 22:21 |
@sonney2k | wiking, or ensure that the randomly generated vector elements are all in range 0...1 | 22:21 |
wiking | should be the same as sum(a) before scaling | 22:21 |
wiking | or? | 22:21 |
blackburn | heiko1: what?? :D | 22:21 |
heiko1 | blackburn, yeah weird :D | 22:21 |
heiko1 | we got a muslim banladesh majority here | 22:22 |
heiko1 | brick land east london | 22:22 |
heiko1 | lane | 22:22 |
blackburn | omg | 22:22 |
blackburn | I thought it is a londongrad but it is a londonesh | 22:23 |
heiko1 | yeah | 22:23 |
heiko1 | especially here | 22:23 |
heiko1 | but it has advantages (free banlgadesh and indian food from time to time :) | 22:23 |
heiko1 | blackburn, I get now why you wanted to get the original indices | 22:24 |
heiko1 | you wanted to search them | 22:24 |
wiking | doh i dont get this shiiiaaat | 22:24 |
blackburn | yeah something like that | 22:24 |
CIA-18 | shogun: Soeren Sonnenburg master * r35d423d / .gitignore : update git ignore to ignore only certain files/dirs in shogun/doc - http://git.io/4DbSiA | 22:24 |
CIA-18 | shogun: Soeren Sonnenburg master * rf606639 / (.gitmodules doc/tutorial): add doc/tutorial as submodule - http://git.io/ZL2iUw | 22:24 |
heiko1 | mmh, but thats actually not really sugsetting anymore | 22:24 |
heiko1 | cant you create a class | 22:24 |
heiko1 | that checks in which tasks an index is? | 22:25 |
heiko1 | like the other way round? | 22:25 |
heiko1 | and then build subsets from that | 22:25 |
wiking | DOH I DONT GET IIIT! | 22:25 |
gsomix | sonney2k, ufff, big, big example. | 22:25 |
wiking | sonney2k heiko1 i've pushed the requested changes... check out the PR | 22:27 |
@sonney2k | wiking, ? relative accuracy is just dividing by some number which is in the order of the scale of the data... but maybe it is better to just put the data on 'right' scale in this test | 22:27 |
wiking | sonney2k: check the commit | 22:28 |
@sonney2k | wiking, url? | 22:28 |
wiking | https://github.com/vigsterkr/shogun/commit/ed46678a6abdef5945449e3621233d63cd269b0a | 22:28 |
shogun-buildbot | build #97 of bsd1 - libshogun is complete: Failure [failed compile] Build details are at http://www.shogun-toolbox.org/buildbot/builders/bsd1%20-%20libshogun/builds/97 blamelist: Soeren Sonnenburg <sonne@debian.org> | 22:29 |
@sonney2k | wiking, which line? | 22:29 |
heiko1 | wiking, ok for me | 22:29 |
gsomix | sonney2k, done | 22:31 |
gsomix | sonney2k, sorry for one big commit. I'm stupid :( | 22:31 |
shogun-buildbot | build #98 of bsd1 - libshogun is complete: Success [build successful] Build details are at http://www.shogun-toolbox.org/buildbot/builders/bsd1%20-%20libshogun/builds/98 | 22:32 |
@sonney2k | gsomix, yes SATAN 666 :D | 22:33 |
@sonney2k | https://github.com/shogun-toolbox/shogun/pull/666 | 22:35 |
@sonney2k | *lolz* | 22:35 |
-!- blackburn [~blackburn@109.226.92.17] has quit [Quit: Leaving.] | 22:38 | |
@sonney2k | gsomix, massiv patch | 22:39 |
gsomix | uh :( | 22:39 |
@sonney2k | gsomix, one thing - could you please in the example return the subsets (this is for the regression tests...) | 22:41 |
wiking | sson it's L66 | 22:42 |
wiking | sonney2k: 66 | 22:42 |
wiking | https://github.com/vigsterkr/shogun/commit/ed46678a6abdef5945449e3621233d63cd269b0a#L0R66 | 22:42 |
-!- blackburn [~blackburn@109.226.92.17] has joined #shogun | 22:46 | |
@sonney2k | wiking, just use a.random(0.0, 1.0); to get sane values | 22:47 |
@sonney2k | wiking, though I am a bit surprised | 22:47 |
@sonney2k | 1024 is not that big... | 22:47 |
wiking | sonney2k: ok i'll change it... let me know what u wanna do on the end with this... submodule or else... | 22:50 |
@sonney2k | wiking, ok I am no longer surprised | 22:50 |
@sonney2k | double mantissa is 52 bits | 22:50 |
@sonney2k | so precision is about 1024**2*1.0/2**52 ~= 2e-10 | 22:51 |
wiking | i've gotta run out now... but going to be back later so i'll read the logs | 22:51 |
@sonney2k | wiking, so use 1e-15 and 1.0 | 22:51 |
wiking | sonney2k: okey | 22:51 |
@sonney2k | wiking, ok cu | 22:51 |
droopy | cheers | 22:52 |
wiking | sonney2k: i'll do that but now i've gotta run | 22:52 |
wiking | bbl within 2 hours | 22:52 |
gsomix | sonney2k, subsets? | 22:52 |
@sonney2k | gsomix, what? | 22:52 |
@sonney2k | ahh | 22:52 |
@sonney2k | slices | 22:52 |
@sonney2k | your somedenseshogunmatrix[:,3] etc stuff | 22:53 |
gsomix | ok, ok | 22:53 |
-!- heiko1 [~heiko@host86-177-114-238.range86-177.btcentralplus.com] has quit [Quit: Leaving.] | 22:54 | |
gsomix | sonney2k, done, it works, btw. | 22:54 |
gsomix | need to check ident problems | 22:55 |
@sonney2k | gsomix, ok do that and then heiko or blackburn can merge | 23:10 |
@sonney2k | I am off for today | 23:10 |
gsomix | yay \o/ | 23:10 |
gsomix | sonney2k, cu | 23:10 |
@sonney2k | gsomix, ahh btw - how much effort would it be to add this access support to SGVector/Matrix ? | 23:11 |
gsomix | sonney2k, not many. just copy/paste macros for densefeatures | 23:12 |
blackburn | sonney2k: why to do that? | 23:13 |
gsomix | but I think, I just should write more common macros | 23:14 |
gsomix | blackburn, for DirectorClasses, I think | 23:15 |
@sonney2k | blackburn, you mean beacause any sgvector etc instance is converted to numpy stuff? | 23:15 |
blackburn | sounds complex | 23:15 |
@sonney2k | yeah | 23:15 |
blackburn | sonney2k: yeah exactly | 23:16 |
@sonney2k | in directors we didn't have this | 23:16 |
@sonney2k | but a real 'sgvector' object | 23:16 |
@sonney2k | and it might be nice to just say x+=y | 23:16 |
* gsomix hates ctrl+z | 23:16 | |
@sonney2k | etc | 23:16 |
gsomix | compiling... | 23:23 |
blackburn | sonney2k: what is xx? | 23:23 |
blackburn | ah | 23:24 |
blackburn | so in directors we have sgvector ok :) | 23:24 |
gsomix | sonney2k, blackburn done | 23:31 |
gsomix | ahhh | 23:33 |
gsomix | return | 23:33 |
gsomix | blackburn, wait | 23:33 |
gsomix | : | 23:33 |
CIA-18 | shogun: Soeren Sonnenburg master * rb2eabfc / (4 files in 3 dirs): Merge pull request #666 from gsomix/buffer_protocol (+5 more commits...) - http://git.io/V5a30w | 23:34 |
gsomix | oops, ok, later | 23:34 |
gsomix | :( | 23:34 |
--- Log closed Tue Jul 24 00:00:17 2012 |
Generated by irclog2html.py 2.10.0 by Marius Gedminas - find it at mg.pov.lt!