--- Log opened Sun Aug 19 00:00:17 2012 | ||
blackburn | sonney2k: I think I will add sparse features with implicit normalization | 00:02 |
---|---|---|
blackburn | and mean -> 0 | 00:03 |
blackburn | costly thing however | 00:03 |
blackburn | [DEBUG] correct=0, total=2150, rejected=0 | 00:13 |
blackburn | omg | 00:13 |
-!- gsomix [~gsomix@178.45.39.145] has quit [Ping timeout: 245 seconds] | 01:04 | |
-!- CIA-21 [~CIA@204.152.223.100] has quit [Remote host closed the connection] | 01:29 | |
shogun-buildbot_ | build #59 of nightly_none is complete: Failure [failed compile] Build details are at http://www.shogun-toolbox.org/buildbot/builders/nightly_none/builds/59 | 03:00 |
shogun-buildbot_ | build #70 of nightly_default is complete: Failure [failed compile] Build details are at http://www.shogun-toolbox.org/buildbot/builders/nightly_default/builds/70 | 03:02 |
shogun-buildbot_ | build #55 of nightly_all is complete: Failure [failed compile] Build details are at http://www.shogun-toolbox.org/buildbot/builders/nightly_all/builds/55 | 03:03 |
shogun-buildbot_ | build #381 of deb1 - libshogun is complete: Success [build successful] Build details are at http://www.shogun-toolbox.org/buildbot/builders/deb1%20-%20libshogun/builds/381 | 04:26 |
-!- emrecelikten [~emre@trir-5d800450.pool.mediaWays.net] has quit [Ping timeout: 244 seconds] | 04:26 | |
-!- emrecelikten [~emre@trir-4d0d87fa.pool.mediaWays.net] has joined #shogun | 04:40 | |
shogun-buildbot_ | build #391 of deb3 - modular_interfaces is complete: Failure [failed compile java_modular] Build details are at http://www.shogun-toolbox.org/buildbot/builders/deb3%20-%20modular_interfaces/builds/391 blamelist: Sergey Lisitsyn <lisitsyn.s.o@gmail.com> | 04:44 |
-!- n4nd0 [~nando@s83-179-44-135.cust.tele2.se] has joined #shogun | 07:43 | |
n4nd0 | good morning | 07:44 |
blackburn | hey | 09:31 |
-!- hoijui [~hoijui@dslb-092-078-182-230.pools.arcor-ip.net] has joined #shogun | 09:44 | |
-!- gsomix [~gsomix@178.45.39.145] has joined #shogun | 10:53 | |
gsomix | hi | 10:53 |
wiking | ho | 10:59 |
blackburn | n4nd0: so what about risks with hm? | 11:54 |
blackburn | can I help you? | 11:54 |
n4nd0 | blackburn: I have to take a look to the PR that wiking did | 11:54 |
blackburn | well he just put risk to so model | 11:54 |
n4nd0 | for what I undestood | 11:54 |
n4nd0 | it should work with that | 11:55 |
n4nd0 | do you remember the name of the PR? | 11:57 |
blackburn | not really | 11:57 |
n4nd0 | has it already been merged? | 11:57 |
n4nd0 | got it | 11:57 |
n4nd0 | https://github.com/shogun-toolbox/shogun/pull/719 | 11:57 |
blackburn | yes sure | 11:57 |
n4nd0 | I am going to issue a new PR soon and later I will check that | 12:00 |
n4nd0 | this is a small one | 12:00 |
n4nd0 | I just added a new Evaluation class, StructuredAccuracy, to compute accuracy for structured problems | 12:00 |
blackburn | good thing to have | 12:01 |
n4nd0 | definitely | 12:02 |
wiking | what's with the risk fuction then? | 12:09 |
wiking | *function? | 12:09 |
wiking | blackburn: merge my last pr please | 12:10 |
wiking | n4nd0: ping me when u have 10 mins to chat | 12:10 |
blackburn | wiking: CIA died - it is already merged | 12:10 |
n4nd0 | wiking: now is good, tell me | 12:10 |
wiking | blackburn: lol ok | 12:11 |
wiking | n4nd0: ok brb in 1 min | 12:11 |
n4nd0 | ok | 12:11 |
wiking | n4nd0: i need to redesign or soemting structuredmodel | 12:11 |
wiking | but i need your input/consent | 12:11 |
wiking | :) | 12:11 |
n4nd0 | ok | 12:11 |
wiking | btw | 12:11 |
wiking | check my last PR | 12:11 |
n4nd0 | I have not really looked at the PR yet | 12:11 |
n4nd0 | give some more time then | 12:12 |
wiking | i've change how PSI is calculated in latent | 12:12 |
wiking | as this is the only way we can support sparse features | 12:12 |
wiking | https://github.com/shogun-toolbox/shogun/pull/741 | 12:12 |
wiking | as the SGVector<float64_t> psi (...) both in Structured and LatentModel | 12:13 |
wiking | could not support sparse features | 12:13 |
wiking | ok i'll be back in 1 min | 12:13 |
n4nd0 | wiking: but those changes are just in LatentModel | 12:14 |
n4nd0 | wiking, blackburn: I am checking PR 719, I thought that the idea was to move the risk function to the StructuredModel rather than to the children | 12:16 |
n4nd0 | as I see there it is not implemented in CStructuredModel but in CMulticlassModel | 12:17 |
blackburn | how can that be implemented in structured model? | 12:17 |
n4nd0 | do you remember that conversation we had at midnight some days ago? | 12:18 |
n4nd0 | I think that the conclusion there was that the risk function was dependent on the argmax and the psi | 12:18 |
wiking | n4nd0: for now yes | 12:18 |
wiking | n4nd0: that risk function only valid for Multiclass case | 12:18 |
wiking | that's why it's in multiclass | 12:18 |
n4nd0 | but apart from that, the definition is genereal | 12:18 |
blackburn | n4nd0: no, subgrad is not general | 12:18 |
wiking | n4nd0: the risk fucntion for instance for SOlatent svm would be different | 12:19 |
n4nd0 | blackburn: wasn't it too in the definition we saw in that thesis? | 12:19 |
n4nd0 | wiking: that's why I think it should be virtual, but in StructuredModel | 12:19 |
blackburn | yes, for some models it is | 12:19 |
shogun-buildbot_ | build #392 of deb3 - modular_interfaces is complete: Failure [failed compile java_modular] Build details are at http://www.shogun-toolbox.org/buildbot/builders/deb3%20-%20modular_interfaces/builds/392 blamelist: Viktor Gal <viktor.gal@maeth.com> | 12:19 |
wiking | n4nd0: and about the psi function: i've just done it in latent model since i don't know how it would effect SO at the moment | 12:19 |
n4nd0 | wiking: ok | 12:19 |
wiking | n4nd0: it is virtual in structuredModel | 12:20 |
n4nd0 | wiking: let me show you the reference we were talking about | 12:20 |
n4nd0 | well I think there is a mail with that | 12:20 |
wiking | n4nd0: https://github.com/shogun-toolbox/shogun/pull/719/files#L18R228 | 12:20 |
blackburn | n4nd0: I don't mind it to be general | 12:20 |
blackburn | but if there is loss of efficiency I don't mind to have some duplicated code | 12:21 |
wiking | somebody broke java modular :) | 12:21 |
wiking | i guess it's heiko | 12:21 |
n4nd0 | http://users.cecs.anu.edu.au/~chteo/pub/Teo10_phdthesis.pdf | 12:21 |
blackburn | I mean if {argmax+subgrad} in multiclass is more efficient than virtual one - we should go this way | 12:22 |
blackburn | see what I mean? | 12:22 |
n4nd0 | blackburn: why should it be more efficient that way? | 12:22 |
wiking | n4nd0: which page? | 12:22 |
blackburn | wiking: Appendix 1 | 12:22 |
n4nd0 | wiking: equations A.1 and A.8, pages 103 and 105 | 12:22 |
n4nd0 | they are the contribution in the risk of one training example and the subgradient, respectively | 12:23 |
wiking | ok | 12:23 |
wiking | but isn't this application specific? | 12:24 |
wiking | i mean what i understood from uricamic's solver | 12:24 |
wiking | it pretty much lambda*||w||_2^2 + Risk | 12:24 |
wiking | i mean it solves this | 12:24 |
blackburn | hahaha I just noticed the author's name | 12:25 |
wiking | and as you can see from the risk function of MulticlassModel... it pretty much works only if your SO problem is a multiclass problem | 12:26 |
blackburn | bad bad word in the middle | 12:26 |
n4nd0 | wiking: aren't the application speficic parts only in the argmax and the psi? | 12:26 |
wiking | n4nd0: in that risk funciton no | 12:26 |
wiking | *function | 12:27 |
blackburn | n4nd0: so you want to implement subgrad in independent fashion as well | 12:27 |
blackburn | right? | 12:27 |
wiking | it loops through the various classes | 12:27 |
n4nd0 | wiking: in uricamic's one? | 12:27 |
wiking | n4nd0: yes | 12:27 |
n4nd0 | wiking: can you point out for me the loops you mean? lines? | 12:27 |
wiking | line 181 | 12:28 |
wiking | in MulticlassModel.cpp | 12:28 |
n4nd0 | wiking: it looks to me that this part is the argmax | 12:28 |
wiking | well if yes then chuck it out :) | 12:29 |
wiking | and replace it with argmax :) | 12:29 |
wiking | and see if we can get further with the story | 12:29 |
n4nd0 | hehe I guess I should read it completely first | 12:29 |
wiking | and move the risk function in structredmodel ;) | 12:29 |
wiking | R({\bf w}) = \sum_{i=1}^{m} \max_{y \in \mathcal{Y}} \left[ \ell(y_i, y) * + \langle {\bf w}, \Psi(x_i, y) - \Psi(x_i, y_i) \rangle \right] | 12:30 |
wiking | this is what the code does apparently | 12:30 |
wiking | (see the doxygen for virtual float64_t risk(float64_t* subgrad, float64_t* W, TMultipleCPinfo* info=0); in MulticlassModel.h | 12:30 |
wiking | but let me know when u r through this one as i have other concerns :) | 12:31 |
n4nd0 | tell me about them :) | 12:32 |
wiking | heheh so ok the following is the story | 12:32 |
wiking | i need to be able to chuck in extra variables for StructuredModel | 12:32 |
wiking | for delta_loss | 12:32 |
wiking | and get_joint_vector... | 12:32 |
wiking | which would mean that i need to change the function arguments in StructuredModel | 12:33 |
wiking | but that won't work :) | 12:33 |
wiking | i mean i need something like this for example for delta loss | 12:33 |
wiking | loss(y, yhat, hhat) instead of loss(y1, y2) | 12:34 |
n4nd0 | ok | 12:34 |
n4nd0 | I thought of that | 12:34 |
wiking | get_joint_feature_vector should be | 12:35 |
n4nd0 | my idea was to put that method in StructuredModel too | 12:35 |
wiking | get_joint_feature_vector(feat_index, structureddata y, latentdata h) | 12:35 |
blackburn | n4nd0: btw about coffin integration | 12:35 |
blackburn | subgrad of models (by Teo) | 12:35 |
blackburn | is 0 + \phi(x,\overline{y}) | 12:36 |
n4nd0 | wiking: just put them in the StructuredModel | 12:36 |
blackburn | - \phi(x,y) | 12:36 |
wiking | n4nd0: how? :) | 12:36 |
blackburn | I mean it is two add_to_dense_vec operations | 12:36 |
wiking | i can do something like: get_joint_feature_vector(feat_index, structureddata y, latentdata h = NULL) | 12:36 |
n4nd0 | wiking: that's why I didn't do delta_loss and get_joint_feature_vector pure virtual, but unimplemented | 12:36 |
wiking | and same for delta_loss | 12:36 |
n4nd0 | exactly, that's the spirit :) | 12:36 |
wiking | ? | 12:36 |
wiking | no comprende | 12:36 |
n4nd0 | mmm | 12:37 |
n4nd0 | what is the problem of making new methods? | 12:37 |
n4nd0 | we have a delta_loss(y1,y2) | 12:37 |
n4nd0 | put another one there delta_loss(y,yhat,hhat) | 12:37 |
blackburn | wiking: can you convert it to sth like 'add_joint_feature_vector_to_dense_vec'? :) | 12:37 |
shogun-buildbot_ | build #393 of deb3 - modular_interfaces is complete: Failure [failed compile java_modular] Build details are at http://www.shogun-toolbox.org/buildbot/builders/deb3%20-%20modular_interfaces/builds/393 blamelist: Sergey Lisitsyn <lisitsyn.s.o@gmail.com> | 12:38 |
wiking | blackburn: which where? | 12:38 |
blackburn | wiking: I mean now you are using get_joint_feature_vector | 12:38 |
wiking | n4nd0: and who will call it in your Primal solver? | 12:38 |
blackburn | oh I will wait you to finish :) | 12:38 |
wiking | blackburn: which class r u talking about | 12:39 |
blackburn | wiking: let me check last pr again | 12:39 |
n4nd0 | wiking: aham, so you would like to use the primal solver for your latent problems directly | 12:40 |
wiking | yes | 12:40 |
wiking | as well as the dual solver by uricamic | 12:40 |
n4nd0 | wiking: aham, are you sure it can be done like that? | 12:40 |
wiking | well the only difference is | 12:40 |
wiking | how you calculate the delta loss and the psi | 12:40 |
wiking | i.e. psi is based on x,y,h | 12:41 |
wiking | and not only on x,y | 12:41 |
wiking | same with the delta loss | 12:41 |
n4nd0 | ok | 12:41 |
wiking | see here: | 12:41 |
n4nd0 | but even if you include a new parameter latentdata h = NULL | 12:41 |
n4nd0 | how do you plan to make the difference between executing for latent or not? | 12:42 |
n4nd0 | some kind of boolean member? | 12:42 |
wiking | n4nd0: no | 12:42 |
wiking | i mean from the fucntion you can tell whether the func arg is set or not | 12:42 |
wiking | right? | 12:42 |
wiking | btw i can see that you use only argmax directly from PrimalMosek | 12:43 |
wiking | right? | 12:43 |
n4nd0 | yes, I think so | 12:43 |
n4nd0 | wiking: and yes, you are right about the thing of the func arg; I am king of retarded... | 12:44 |
n4nd0 | kind* of | 12:44 |
wiking | mmm | 12:44 |
wiking | then maybe this would work | 12:44 |
wiking | but the stupid thing is | 12:44 |
wiking | mmm | 12:44 |
wiking | somehow i wanted to hide this behind the function if you know what i mean | 12:45 |
wiking | so that things are not passed via the function arg | 12:45 |
wiking | rather you just give an index or something | 12:45 |
wiking | and everything else is behind the scenes | 12:45 |
wiking | blackburn: ok check it again | 12:46 |
wiking | n4nd0: but maybe it'll work what i'm thinking of now | 12:46 |
blackburn | check what? | 12:46 |
wiking | as really delta_loss and get_joint_feature_vector is being called in your argmax | 12:46 |
wiking | blackburn: last pr | 12:46 |
wiking | blackburn: you were about to comment there something | 12:47 |
blackburn | I'm kind of lost there | 12:47 |
wiking | blackburn: this one: https://github.com/shogun-toolbox/shogun/pull/741 | 12:48 |
wiking | ? | 12:48 |
wiking | in this one? | 12:48 |
blackburn | no in general | 12:48 |
blackburn | so you compute dense features | 12:48 |
blackburn | according to position of window | 12:48 |
blackburn | right? | 12:48 |
wiking | blackburn: mmm don't mix it now with the application :))) | 12:48 |
blackburn | but is it so? | 12:49 |
wiking | blackburn: yeah there is so like that | 12:49 |
wiking | *it is like that | 12:49 |
blackburn | it sounds like memory eating thing | 12:49 |
n4nd0 | wiking: ok, so tell me if so when you have ready with what that you were thinking of | 12:50 |
wiking | n4nd0: well now i'm thinking of creating a class that is actually a direct inheritance of StructuredModel | 12:50 |
wiking | n4nd0: and add there non-virtual functions for delta_loss and get_psi | 12:51 |
wiking | n4nd0: and the argmax and basically then the argmax will call the right functions with the right amount of arguments | 12:51 |
wiking | and then for now i don't touch the StructuredModel | 12:51 |
wiking | and if that one works out then we'll see later about how we could simplify the thing | 12:52 |
wiking | i have no better idea now :S | 12:52 |
n4nd0 | it looks like a good idea to me | 12:52 |
n4nd0 | probably those two classes could be merged once we check it works | 12:52 |
n4nd0 | there the PR goes :) | 13:31 |
blackburn | ?????? | 13:43 |
blackburn | merged | 13:43 |
n4nd0 | thank you :) | 13:44 |
wiking | cia bot die! | 13:58 |
wiking | and it did apparently | 13:59 |
shogun-buildbot_ | build #394 of deb3 - modular_interfaces is complete: Failure [failed compile java_modular] Build details are at http://www.shogun-toolbox.org/buildbot/builders/deb3%20-%20modular_interfaces/builds/394 blamelist: iglesias <fernando.iglesiasg@gmail.com> | 14:08 |
blackburn | n4nd0: did you read about CRFs before? is it's argmax hard to implement? | 14:27 |
n4nd0 | blackburn: I have read something about CRFs ues | 14:28 |
n4nd0 | yes* | 14:28 |
n4nd0 | I don't think the argmax is something fancier than Viterbi | 14:28 |
blackburn | I was pretty excited when I saw CRF based segmentation | 14:29 |
n4nd0 | BUT the thing is that I don't know how CRFs would fit into our SO framework | 14:29 |
n4nd0 | why? | 14:29 |
blackburn | why not | 14:29 |
blackburn | ? | 14:29 |
blackburn | what is so different between HM and CRFs? | 14:29 |
n4nd0 | in the paper I read the way of training had nothing to do with the one for HM-SVM I know | 14:30 |
blackburn | hmm I see | 14:32 |
shogun-buildbot_ | build #395 of deb3 - modular_interfaces is complete: Failure [failed compile java_modular] Build details are at http://www.shogun-toolbox.org/buildbot/builders/deb3%20-%20modular_interfaces/builds/395 blamelist: iglesias <fernando.iglesiasg@gmail.com>, Sergey Lisitsyn <lisitsyn.s.o@gmail.com> | 14:33 |
n4nd0 | blackburn: the build was already broken before right? | 14:34 |
blackburn | yes | 14:34 |
blackburn | by heiko | 14:34 |
wiking | neeeeeeedcoffee | 14:49 |
-!- heiko [~heiko@host86-185-9-87.range86-185.btcentralplus.com] has joined #shogun | 14:50 | |
wiking | heiko: broker! :) wazza? | 15:04 |
heiko | wiking hello | 15:04 |
heiko | ehm | 15:05 |
heiko | I didn't do it :D | 15:05 |
wiking | CRF CRF blackburn will u implement? :) | 15:05 |
-!- pluskid [~pluskid@76.79.79.246] has joined #shogun | 15:15 | |
shogun-buildbot_ | build #396 of deb3 - modular_interfaces is complete: Failure [failed compile java_modular] Build details are at http://www.shogun-toolbox.org/buildbot/builders/deb3%20-%20modular_interfaces/builds/396 blamelist: Heiko Strathmann <heiko.strathmann@gmail.com> | 15:34 |
-!- pluskid [~pluskid@76.79.79.246] has quit [Ping timeout: 276 seconds] | 15:34 | |
heiko | shogun-buildot_ status | 15:49 |
heiko | wiking! | 15:51 |
heiko | unit-tests! | 15:51 |
heiko | still dont work | 15:51 |
heiko | gtest refuses to install | 15:51 |
heiko | how to I do this? | 15:51 |
shogun-buildbot_ | build #397 of deb3 - modular_interfaces is complete: Failure [failed compile java_modular] Build details are at http://www.shogun-toolbox.org/buildbot/builders/deb3%20-%20modular_interfaces/builds/397 blamelist: Heiko Strathmann <heiko.strathmann@gmail.com> | 15:52 |
wiking | heiko: ahahhaha yeah i told u to ask blackburn | 16:00 |
wiking | heiko: i dont know wtf is with that shit but the best is that you do the following | 16:01 |
wiking | sudo cp lib/.libs/libgtest.so* /usr/local/lib | 16:01 |
wiking | sudo cp lib/.libs/libgtest_main.so* /usr/local/lib | 16:01 |
wiking | heiko: and then the configure should work finally | 16:01 |
wiking | i dont know why do they do this at google... i mean to refuse to do a make install :) | 16:03 |
shogun-buildbot_ | build #398 of deb3 - modular_interfaces is complete: Failure [failed compile java_modular] Build details are at http://www.shogun-toolbox.org/buildbot/builders/deb3%20-%20modular_interfaces/builds/398 blamelist: Heiko Strathmann <heiko.strathmann@gmail.com> | 16:05 |
-!- hoijui [~hoijui@dslb-092-078-182-230.pools.arcor-ip.net] has quit [Ping timeout: 265 seconds] | 16:16 | |
shogun-buildbot_ | build #399 of deb3 - modular_interfaces is complete: Failure [failed compile java_modular] Build details are at http://www.shogun-toolbox.org/buildbot/builders/deb3%20-%20modular_interfaces/builds/399 blamelist: Heiko Strathmann <heiko.strathmann@gmail.com> | 16:23 |
-!- hoijui [~hoijui@dslb-092-078-182-230.pools.arcor-ip.net] has joined #shogun | 16:26 | |
-!- hoijui [~hoijui@dslb-092-078-182-230.pools.arcor-ip.net] has quit [Ping timeout: 260 seconds] | 16:33 | |
n4nd0 | /exit | 17:14 |
-!- n4nd0 [~nando@s83-179-44-135.cust.tele2.se] has quit [Quit: leaving] | 17:14 | |
-!- wubino [~wubino@74.61.91.195] has joined #shogun | 17:47 | |
-!- puffin444 [180317bb@gateway/web/freenode/ip.24.3.23.187] has joined #shogun | 17:58 | |
puffin444 | hey blackburn | 17:59 |
-!- heiko [~heiko@host86-185-9-87.range86-185.btcentralplus.com] has left #shogun [] | 18:18 | |
puffin444 | hey blackburn, just to let you know, i rebased my repo so its ready to be merged | 18:20 |
blackburn | puffin444: hey | 18:23 |
blackburn | okay nice | 18:23 |
-!- heiko [~heiko@host86-185-9-87.range86-185.btcentralplus.com] has joined #shogun | 18:32 | |
-!- heiko [~heiko@host86-185-9-87.range86-185.btcentralplus.com] has quit [Client Quit] | 18:35 | |
blackburn | wiking: I will if n4ndo not | 18:40 |
shogun-buildbot_ | build #289 of cyg1 - libshogun is complete: Failure [failed compile] Build details are at http://www.shogun-toolbox.org/buildbot/builders/cyg1%20-%20libshogun/builds/289 blamelist: Heiko Strathmann <heiko.strathmann@gmail.com> | 18:42 |
-!- wubino [~wubino@74.61.91.195] has left #shogun ["Leaving"] | 19:23 | |
-!- hoijui [~hoijui@dslb-092-078-182-230.pools.arcor-ip.net] has joined #shogun | 19:50 | |
wiking | blackburn: :D | 20:58 |
-!- os253 [~os@e180116168.adsl.alicedsl.de] has joined #shogun | 20:59 | |
os253 | Hi puffin444 | 21:00 |
puffin444 | hello os253 | 21:00 |
os253 | is now good for you?I am around all evening. | 21:01 |
puffin444 | yes now is good if its good for you | 21:01 |
blackburn | oh gaussian team in da house | 21:01 |
puffin444 | yes indeed :) | 21:02 |
os253 | yes, fine here. | 21:02 |
os253 | :-) | 21:02 |
blackburn | tell me - how much effort is to add GP classifiers? | 21:02 |
puffin444 | honestly not sure. I would have to check GPML. | 21:02 |
puffin444 | Shouldnt require too much though | 21:03 |
os253 | How far is the Laplace code? If that works it's just another likelihood function... | 21:03 |
puffin444 | The laplace code is pretty much done and ready to be merged wt any time. | 21:03 |
blackburn | about your PR | 21:04 |
puffin444 | yes? | 21:04 |
blackburn | puffin444: you merged it in some kind of broken way - can you clean it up? | 21:04 |
puffin444 | uh oh what's the problem? | 21:05 |
blackburn | os253: have you seen puffin444 PR? https://github.com/shogun-toolbox/shogun/pull/707 | 21:05 |
blackburn | puffin444: latest commits include some heiko's ones.. | 21:05 |
blackburn | however may be it is not worth it | 21:05 |
os253 | Then it should be easy. blackburn: do you think classification is useful? Generaly speaking: gps are not that great for classifcation or at least not better than SVMs. | 21:05 |
blackburn | os253: I am pretty lame in everything about GPs :) | 21:06 |
blackburn | os253: I am kind of chris widmer representer :D | 21:07 |
os253 | i see | 21:07 |
puffin444 | os253 all the main features mentioned in the project plan have been merged or are ready to be merged | 21:09 |
blackburn | puffin444: can you change namespace in brent method source? | 21:09 |
puffin444 | At this point any additional feature should be a fairly easy task | 21:09 |
blackburn | puffin444: I have two general issues I think would be cool to solve | 21:11 |
blackburn | first is rather mandatory - FITC eigen members | 21:11 |
puffin444 | yes? | 21:11 |
blackburn | and second is some alternative to nlopt in gradient model selection | 21:11 |
blackburn | no idea if you have time for that :) | 21:11 |
puffin444 | Well gradient model selection is pretty general | 21:12 |
blackburn | eigen member is serious issue we need to resolve before release | 21:12 |
blackburn | members* | 21:13 |
puffin444 | What about FITC eigen membes? Is that the map<vecrorxd> we discussed earlier? | 21:13 |
-!- n4nd0 [~nando@s83-179-44-135.cust.tele2.se] has joined #shogun | 21:13 | |
blackburn | yes | 21:13 |
blackburn | so please focus on that before 1st of sept | 21:13 |
puffin444 | okay. So replace those Vectorxd members with map? | 21:13 |
blackburn | no, replace it with SGMatrix and SGVector | 21:14 |
blackburn | and map in your code | 21:14 |
os253 | what is the issue with eigen members? | 21:15 |
blackburn | os253: eigen should not be included and/or exposed in headers to avoid additional dependency | 21:15 |
puffin444 | in .h or also .cpp? | 21:15 |
blackburn | puffin444: I think I have an idea | 21:15 |
blackburn | I am going to add as_eigen method | 21:15 |
blackburn | to both sgvector and sgmatrix | 21:16 |
os253 | I see | 21:16 |
blackburn | which returns eigen analogue of the matrix | 21:16 |
blackburn | or the vector | 21:16 |
blackburn | see what I mean? | 21:16 |
puffin444 | yes. | 21:16 |
blackburn | oops not possible | 21:16 |
blackburn | same reason | 21:17 |
blackburn | :D | 21:17 |
blackburn | I missed that | 21:17 |
blackburn | you have to map it manually | 21:17 |
blackburn | okay just change it to sg* and map in your code before use | 21:17 |
puffin444 | Okay. Can I still use eigen vectors directly in cpp files? | 21:17 |
blackburn | yes sure | 21:18 |
puffin444 | os253 I am working ona tutorial pagefor gps | 21:19 |
os253 | yes, that would be great. I have looked through the PR yesterday. It seems like a good mix of sticking to the GPML code and adapting it to your framework. | 21:20 |
blackburn | puffin444: Map<MatrixXd> (m_some_matrix.matrix,m_some_matrix.num_rows,m_some_matrix.num_cols); | 21:20 |
blackburn | that's what you would have to do | 21:20 |
os253 | I am wondering how to best make sure it will be picked up. What examples do you currently have? | 21:20 |
blackburn | err missed name | 21:20 |
os253 | It looks like tutorials and examples are the main thing to add. | 21:20 |
puffin444 | I have examples for every msjor feature under libshogun | 21:20 |
os253 | Perhaps the | 21:20 |
puffin444 | I also have a basic graphical example in python | 21:21 |
puffin444 | All the main features have been integrated into the modular interface | 21:21 |
os253 | I'd vote for a python example, just a single one where you use standard GP, GP with ficti (also visusalizing the location of the learnt inducing points). | 21:22 |
blackburn | puffin444: btw are you going to research GPs further? | 21:22 |
os253 | Lastly the same with student-T. Something like that. | 21:23 |
puffin444 | for graduate school? | 21:23 |
blackburn | no idea, in general | 21:23 |
puffin444 | I hope to stay and contribute to shogun and flesh the gp code out a bit | 21:24 |
blackburn | cool | 21:24 |
puffin444 | os253 that should beeasyto do. | 21:24 |
os253 | puffin444: you almost should stay on, having coded up this chunk is substantial. From that basis it would be easy to carry out novel directions that are more research like. | 21:25 |
os253 | Great, I think a visual python example would be very useful to have. | 21:25 |
puffin444 | Yes I could just build onto the current one I have. | 21:26 |
puffin444 | blackburn, is it possible that the current pr be merged? The only reason I ask is that I have to fix a ton of conflicts each time I have to rebase. | 21:29 |
os253 | that would be great. | 21:29 |
blackburn | puffin444: ah sure I forgot | 21:29 |
blackburn | merged | 21:30 |
blackburn | CIA died so no message | 21:30 |
puffin444 | os253, Do you have any orther questions/concerns? | 21:31 |
os253 | I have a list of small things, but perhaps I can just email them. For example: choice of the number of inducing points. | 21:32 |
os253 | You currently take the same approach than GPML, right? I.e. they are fixed. | 21:34 |
puffin444 | fixed at what point? They are chosen before hyperparameter learning, | 21:35 |
os253 | Or do you have code for choosing an apprpriate number? A simple heuristic may be good enough. | 21:35 |
puffin444 | Oh yes in that case they are fixed. | 21:35 |
shogun-buildbot_ | build #389 of deb1 - libshogun is complete: Failure [failed compile] Build details are at http://www.shogun-toolbox.org/buildbot/builders/deb1%20-%20libshogun/builds/389 blamelist: puffin444 <walke434@msu.edu> | 21:36 |
os253 | Perhaps one could add this to a python demo for large data. Do you have a suitable datasets with thousands of data points to illustrate the use of the sparse approximation? | 21:36 |
shogun-buildbot_ | build #390 of deb1 - libshogun is complete: Failure [failed compile] Build details are at http://www.shogun-toolbox.org/buildbot/builders/deb1%20-%20libshogun/builds/390 blamelist: puffin444 <walke434@msu.edu>, root <root@gs12672-vm.sp.cs.cmu.edu>, Sergey Lisitsyn <lisitsyn.s.o@gmail.com>, puffin444 <walke434@gmail.com> | 21:36 |
puffin444 | No I do not. | 21:37 |
os253 | What datasets are you using at the moment? | 21:38 |
puffin444 | A simple 4 point dataset I use to compare with gpml. | 21:38 |
os253 | Ok. I can send you some suggestion. | 21:39 |
puffin444 | I did this to ensure the hyperparameter learning came up with consistent parameters across gpml and shogun. | 21:39 |
puffin444 | yes. | 21:39 |
os253 | Perhaps best to pick some one can download from a script to avoid having to add it to the repository. | 21:40 |
n4nd0 | blackburn: hey there | 21:40 |
os253 | It would be useful to have a script that can be run, downloads the datasets and processes it alongside some plots. | 21:40 |
n4nd0 | blackburn: when someone publishes a paper, don't you think that the code and data they have used for their results should be provided? | 21:41 |
os253 | I'll send you a list of datasets that are public and can be downloaded. | 21:41 |
puffin444 | I could try that. | 21:41 |
n4nd0 | blackburn: I am quite surprised that I don't manage to find anything about it for the papers I am reading, maybe I am searching wrong :S | 21:41 |
puffin444 | thanks | 21:41 |
-!- gsomix_ [~gsomix@178.45.41.90] has joined #shogun | 21:42 | |
os253 | Yes, that would be great. Thanks. I'll send this by tomorrow. | 21:43 |
-!- gsomix [~gsomix@178.45.39.145] has quit [Ping timeout: 244 seconds] | 21:45 | |
os253 | If you are around tomorrow we can briefly chat then. | 21:46 |
os253 | I am online all day but not necessarily on IRC. | 21:46 |
puffin444 | sorry blackburn. It looks like I forgot a HAVE_EIGEN somewhere | 21:46 |
blackburn | n4nd0: yeah kind of open source science | 21:46 |
blackburn | I believe most cases of not available code means NO CODE | 21:47 |
blackburn | and no research | 21:47 |
n4nd0 | blackburn: but how can it be possible that there is no code? | 21:47 |
blackburn | I mean it could turn to be a false research | 21:47 |
n4nd0 | I mean, the results that are published must be checked somehow! | 21:47 |
blackburn | :D | 21:47 |
puffin444 | os253, I unfortunately will not be available tomorrow. | 21:47 |
blackburn | no they are not checked by editors | 21:47 |
n4nd0 | blackburn: really? | 21:47 |
blackburn | n4nd0: well they are checked for consistency or so | 21:48 |
blackburn | but nobody will try to run your code | 21:48 |
os253 | puffin444: ok -then let's do via email and or at a later time. | 21:48 |
puffin444 | ok. | 21:49 |
n4nd0 | blackburn: I am really surprised | 21:49 |
n4nd0 | when I start writing papers | 21:49 |
n4nd0 | I will offer a link to the code! | 21:49 |
n4nd0 | hehe | 21:49 |
blackburn | yeah that's how it should be | 21:50 |
blackburn | puffin444: can you fix it? | 21:50 |
puffin444 | I will try. | 21:50 |
blackburn | okay thanks | 21:50 |
puffin444 | Can't do it immediately, but i can to it in a few hours | 21:51 |
blackburn | I will do then | 21:51 |
blackburn | puffin444: oh.. you added new methods with eigen return type | 21:52 |
puffin444 | uh oh. How can i resolve it? | 21:53 |
blackburn | the same - return sgvectors/sgmatrices | 21:54 |
blackburn | and then map | 21:54 |
puffin444 | ok. will do. | 21:55 |
blackburn | I wish we had some solution for that but I think depending on eigen is wrong | 21:55 |
puffin444 | ok. | 21:56 |
os253 | blackburn: the more code requirew | 21:57 |
os253 | *requires Eigen the more dominant implicit dependencies become anyway. At some point it will make little difference :-) | 21:58 |
blackburn | os253: yeah that's clear | 21:58 |
blackburn | I have mixed point on that.. | 21:58 |
blackburn | at some point we won't be able to avoid that | 21:59 |
os253 | true... | 22:01 |
shogun-buildbot_ | build #290 of cyg1 - libshogun is complete: Failure [failed compile] Build details are at http://www.shogun-toolbox.org/buildbot/builders/cyg1%20-%20libshogun/builds/290 blamelist: Heiko Strathmann <heiko.strathmann@gmail.com> | 22:03 |
blackburn | shogun-buildbot_: wait a minute wait a minute | 22:09 |
puffin444 | blackburn, should I try to fix the buid in a few hours or are you doing it? | 22:09 |
blackburn | puffin444: I am doing | 22:10 |
puffin444 | ok. | 22:10 |
puffin444 | Then i will get to work on the mao stuff | 22:10 |
puffin444 | map* | 22:10 |
shogun-buildbot_ | build #391 of deb1 - libshogun is complete: Failure [failed compile] Build details are at http://www.shogun-toolbox.org/buildbot/builders/deb1%20-%20libshogun/builds/391 blamelist: Sergey Lisitsyn <lisitsyn.s.o@gmail.com> | 22:20 |
n4nd0 | good night people | 22:43 |
-!- n4nd0 [~nando@s83-179-44-135.cust.tele2.se] has quit [Quit: leaving] | 22:44 | |
-!- os253 [~os@e180116168.adsl.alicedsl.de] has quit [Remote host closed the connection] | 22:52 | |
-!- blackburn [~blackburn@62.106.106.114] has quit [Quit: Leaving.] | 23:01 | |
-!- blackburn [~blackburn@62.106.106.114] has joined #shogun | 23:02 | |
-!- blackburn1 [~blackburn@31.28.44.3] has joined #shogun | 23:05 | |
-!- blackburn [~blackburn@62.106.106.114] has quit [Ping timeout: 252 seconds] | 23:06 | |
-!- blackburn2 [~blackburn@109.226.112.41] has joined #shogun | 23:06 | |
-!- CIA-36 [cia@198.71.88.9] has joined #shogun | 23:09 | |
-!- blackburn1 [~blackburn@31.28.44.3] has quit [Ping timeout: 240 seconds] | 23:09 | |
shogun-buildbot_ | build #392 of deb1 - libshogun is complete: Failure [failed compile] Build details are at http://www.shogun-toolbox.org/buildbot/builders/deb1%20-%20libshogun/builds/392 blamelist: Sergey Lisitsyn <lisitsyn.s.o@gmail.com> | 23:15 |
shogun-buildbot_ | build #393 of deb1 - libshogun is complete: Failure [failed compile] Build details are at http://www.shogun-toolbox.org/buildbot/builders/deb1%20-%20libshogun/builds/393 blamelist: Sergey Lisitsyn <lisitsyn.s.o@gmail.com> | 23:20 |
blackburn2 | mother fucker | 23:26 |
blackburn2 | will that end? | 23:26 |
blackburn2 | shogun-buildbot_: you are killing me bastard | 23:35 |
blackburn2 | puffin444: I have to ifdef just everything | 23:37 |
shogun-buildbot_ | build #394 of deb1 - libshogun is complete: Success [build successful] Build details are at http://www.shogun-toolbox.org/buildbot/builders/deb1%20-%20libshogun/builds/394 | 23:45 |
shogun-buildbot_ | build #315 of deb2 - static_interfaces is complete: Failure [failed test libshogun] Build details are at http://www.shogun-toolbox.org/buildbot/builders/deb2%20-%20static_interfaces/builds/315 blamelist: Sergey Lisitsyn <lisitsyn.s.o@gmail.com> | 23:48 |
shogun-buildbot_ | build #291 of bsd1 - libshogun is complete: Failure [failed test] Build details are at http://www.shogun-toolbox.org/buildbot/builders/bsd1%20-%20libshogun/builds/291 blamelist: Sergey Lisitsyn <lisitsyn.s.o@gmail.com> | 23:51 |
--- Log closed Mon Aug 20 00:00:17 2012 |
Generated by irclog2html.py 2.10.0 by Marius Gedminas - find it at mg.pov.lt!