Open in new window / Try shogun cloud
--- Log opened Tue Oct 02 00:00:17 2012
Guest88299I'm trying to run shogun 2.0.0 with code and data that run fine under shogun 1.1.000:24
Guest88299but I get this error:TypeError: Cannot create new instances of type 'Labels'00:25
Guest88299I think this is the samep problem that is mentioned here: - but is unsolved00:25
Guest88299any ideas?00:25
shogun-buildbotbuild #120 of nightly_default is complete: Failure [failed test]  Build details are at
-!- Netsplit *.net <-> *.split quits: Guest8829910:22
-!- blackburn [5bdfb203@gateway/web/freenode/ip.] has joined #shogun14:14
wikingblackburn: yo16:19
blackburnwiking: hey16:19
wikingblackburn: u were looking for me16:19
blackburnthese guys from company I am applying to want me to analyze
blackburnI'd appreciate if you share any thoughts :)16:20
wikingthey've opensourced it?16:21
wikingit's just the service16:21
blackburnheh right16:21
blackburnthat makes the task a task16:22
wikingit's quite nice16:22
wikinghave you tried to upload images?16:22
blackburnyes, quite a lot already16:22
wikinghow is it?16:22
wikingbtw: it's fun to upload dummy images :D16:23
blackburnwell I put some blender images16:23
blackburnand smoothed/inverted/colored images, etc16:23
wikingi'm trying this
wikingbut it says there's high traffic now16:24
blackburnI tried green plane16:24
blackburnwanted it to think it is a grass16:24
blackburnno category at all16:24
wikingmmm that's good16:24
blackburncheck ILSVRC 201116:24
blackburnthese guys won it :)16:24
blackburnprobably it is the same system16:25
wikingyeah i know their work16:25
wikingmostly based on fisher vector16:25
blackburnproduct quantization16:25
blackburnlinear svm16:25
blackburnand fisher vectors16:25
blackburnSIFT + color histograms16:25
wikingyep yep16:25
wikingi know that one16:25
blackburndo you know how can I prove that it is a sift?16:26
wikingcsurka is leading their group16:26
wikingmmm good question16:26
blackburnsome computer vision hacking16:26
blackburnI was thinking about it16:26
blackburnbut nothing except blurring16:26
wikingwell yeah u should trick sift16:26
wikingwell what could be actually a funny thing to do16:27
wikingis to seed a random image16:27
wikingwith some patches/sift keypoints of a chicken picture16:27
blackburnI want to backtrace feature building16:27
wikingi bet that's possible16:27
blackburnI thought of skating horse16:27
wikingand that the image u construct is still not a chicken :)16:27
blackburnI will do that tomorrow16:27
blackburnwell but still16:28
blackburnI tried same thing actually16:28
blackburnyesterday I took a hockey player image16:28
blackburnand just 'cut' his legs16:28
blackburn'skating' category disappeared16:28
blackburnproves that it is a local feature thing16:29
wikingso you think puting the skaters on a horse16:29
wikingwould do some trick :)16:29
wikingas said16:29
blackburnyeah horse on skates would make these guys laugh at least16:29
wikingtake a simple sift extractor16:29
wikingand get some patches16:30
wikingand randomly distribute that in a dummy image16:30
blackburnwhy do you think it proves the SIFTness?16:30
wikingand then maybe get 2 sets of differetn sift features (from 2 different image categories) and start randomly mixing them in a new dummy image16:31
blackburnI am thinking about something very special about SIFT16:31
blackburnblurring is one thing - gradients go worse with it16:31
wikingone good thing is an affine transform of the image16:32
wikingto check what type of sift they are using16:32
wikingthe default sift implementation is not so good with affine transformation16:32
blackburnhmm right right16:32
blackburnrotations is what I didn't try yet16:32
blackburngood idea16:32
wikingwell not only rotate16:33
wikingbut like rotate and make a trapezoid of the original image16:33
wikingfor example with the eiffel tower16:33
blackburnwell I'll try some fisheye16:33
blackburnto see whether it is robust16:33
blackburnwiking: do you understand what is a way to handle multiple patches but within the same dimension of feature space?16:51
blackburnthat's the thing I miss16:51
wikingblackburn: eh?16:53
wikingblackburn: i guess they could use some pca16:53
wikingas they usually do in their system at xerox16:53
blackburnI mean if one image has 3 patches16:54
blackburnand other has 916:54
blackburnhow pca could help there?16:54
wikingwell they don't give a damnfuck about the number of patches16:54
wikingas they calculate in a way a frequency vector16:54
wikingof the patches16:54
wikingof the number of various patches16:55
wikingthat is then normalised16:55
wikingso that's your feature vector, i.e. the fisher vector16:55
wikingin the case of xerox16:55
blackburnso # of features is just ~ to # of possible visual words?16:55
wikingyep yep16:55
wikingusual drill16:55
blackburngot it16:55
wikingthe fisher vector is something like a bag of visual words16:55
wikinga bit more scientific16:55
wikingbut on the end it is basically doing that16:56
blackburnmore clear now16:56
wikingheheh welcome :)16:56
wikinghack the shit out of their system :))16:56
wikingif i'd have time i'd really help16:56
blackburnyeah, I should try to hack out sift16:56
wikingbecause this way you could prove some funky security stuff16:56
blackburnno problem you helped a lot already16:56
wikingor let say16:56
wikinghow to SEO and image search engine :P16:57
wikingafaik this has been not really exploited yet16:57
wikingbut for sure it can be16:57
wikingas all the keypoint extractors16:57
wikingare very very16:57
wikingso they must have several flaws16:57
blackburnwiking: what are you working on btw?17:06
-!- blackburn [5bdfb203@gateway/web/freenode/ip.] has quit [Quit: Page closed]17:25
-!- rebecca [a9e64c04@gateway/web/freenode/ip.] has joined #shogun17:41
-!- rebecca is now known as Guest4499117:42
-!- Guest44991 [a9e64c04@gateway/web/freenode/ip.] has quit [Quit: Page closed]17:49
-!- RebeccaTruty [a9e64c04@gateway/web/freenode/ip.] has joined #shogun17:50
RebeccaTrutyhey - anyone there that can answer a question?17:51
RebeccaTrutythis is "Guest88299" from overnight17:51
-!- n4nd0 [] has joined #shogun18:02
wikingRebeccaTruty: hey18:55
wikingRebeccaTruty: can u send the code maybe somewhere? like pastebin18:56
-!- blackburn [~blackburn@] has joined #shogun19:17
wikingblackburn: hahhaha what are the tags?20:02
blackburnwiking: skating, winter sports, donkey20:02
blackburndonkey is the first for sure20:02
blackburnbut skating is here20:02
blackburnfeel myself a hacker20:02
wikingblackburn the hax0r20:03
-!- n4nd0 [] has quit [Ping timeout: 246 seconds]20:05
wikinghow's n4nd0's stuff with pthread? :)20:10
-!- blackburn [~blackburn@] has quit [Quit: Leaving.]20:17
-!- blackburn [~blackburn@] has joined #shogun20:18
RebeccaTrutywiking: I put some code here20:53
-!- n4nd0 [] has joined #shogun20:53
-!- n4nd0 [] has quit [Quit: leaving]21:17
wikingRebeccaTruty: where?21:34
RebeccaTrutyoh sorry, here:
wikingRebeccaTruty: oh yeah21:39
wikingthe thing is that Labels class has been majorly redesign21:39
wikingso you'll have to change your code21:39
RebeccaTrutyI see21:39
wikingso in this case21:40
wikinguse DenseLabels21:40
wikingrandom.choice([0.0, 1.0])21:40
wikinggives you 0 or 1 or a random number in [0,1] interval?21:40
RebeccaTrutyit give you a 1 or 021:40
wikingah ok21:41
wikingthen just use21:41
RebeccaTrutygreat, thanks21:41
wikingthat's the doxygen for the class21:41
wikingbut yeah it shoudl work21:42
wikingif you just replace21:42
wikingtest_labels = Labels(np.array(raw_labels, dtype=np.double))21:42
RebeccaTrutyyup, thanks, sorry I didn't realize it was such a stupid question - I shoudl have looked into it more21:42
wikingtest_labels = BinaryLabels(np.array(raw_labels, dtype=np.double))21:42
wikingRebeccaTruty: na no worries21:42
wikingwe just needed a bit more complex labeling system... :P21:42
RebeccaTrutyunderstood.  thanks for the help!21:42
wikingno worries21:43
wikingbut next time i think you'd get a faster answer on the mailinglist of shogun21:43
wikingmmm fuck where's the link on the new website to the mailing list :)21:44
wikingsonney2k blackburn where's the mailing list information on the new website?21:44
wikingah ok i've found it21:45
blackburnuh you were having kinda discussion21:45
wikingblackburn: ehhehe no worries forget that i was messaging :P21:45
blackburnsomebody asked about labels thing yesterday21:46
blackburnwas that you, RebeccaTruty21:46
RebeccaTrutyyes - that was me21:46
blackburnclear, I am sorry you've got to handle this transformation manually21:46
blackburnwe came to case where we had no other solution because of design gap21:47
RebeccaTrutyyeah, no problem, it should be an easy enough fix.21:48
blackburnwiking: what's that/22:13
wikingstargazers :D22:13
blackburnmy city is not highlighted, is that good? :)22:14
blackburnahhh who stars his project22:14
blackburnwiking: find me!22:15
blackburnthere are only 5 in russia22:15
blackburnwiking: sad no one from north korea22:18
-!- RebeccaTruty [a9e64c04@gateway/web/freenode/ip.] has left #shogun []22:42
wikingthis guy was crazy22:42
blackburnwiking: heheh he has good chiefs22:46
blackburnberia and yezhov22:46
blackburnwiking: did you know that once beria was killed gov't sent an order to all libraries22:49
blackburnto cut down beria article in encyclopedia22:49
blackburnand replace it with beringov proliv22:49
-!- blackburn [~blackburn@] has quit [Quit: Leaving.]23:01
--- Log closed Wed Oct 03 00:00:17 2012