On 08/22/2013 02:38 PM, Björn Esser wrote: > Am Donnerstag, den 22.08.2013, 14:22 +0200 schrieb Lars Buitinck: >> 2013/8/22 Olivier Grisel <[email protected]>: >>> 2013/8/22 Björn Esser <[email protected]>: >>>> Can you please tell me a bit more about what / why these were modded? >>>> If we cannot unbundle them, I need to have some further infos to get >>>> some exception granted for them. :) >>> For the libsvm binding the main reason it to get both the sparse and >>> dense memory layout for the input data. The upstream package only >>> supports sparse input data and the dense variant lives in a 3rd party >>> fork: >> [snip] >>> For liblinear I am not sure anymore (I was not personally involved in >>> those bindings). It's probably worth exploring with diff of the source >>> folders. >> We also added RNG seeding to both LibSVM and Liblinear, so the API and >> calling interface are different from upstream. We also fixed a couple >> of bugs and swapped +1 and -1 in some places to match our conventions >> better. The RNG seeding is important for reproducible results (e.g., >> to make the tests pass reliably). > Not a hint of that in liblinear code... :) Sorry, did you mean there is not a hint of swapping -1 and +1? There is in 0fbf223d7e3dfd43d5eac954f06b989436da8e7c and 6fd64d48d04e06aac43c3f998b00303a24d934d9. The setting of the random seed is done in a helper file, liblinear_helper.c
>> The main changes to Liblinear can be seen with gitk >> sklearn/svm/src/liblinear/linear.cpp in the scikit-learn source >> directory. We have 1.91, btw. > There are ~30 lines added and ~5 lines reordered as I see. There is basically some code to sort the labels added and the C deallocation bug fixed. So you are right, these are not major changes. Building a separate library for this seems quite hassle, though. But you are the one knowing about the fedora guidelines, not me ;) Cheers, Andy ------------------------------------------------------------------------------ Introducing Performance Central, a new site from SourceForge and AppDynamics. Performance Central is your source for news, insights, analysis and resources for efficient Application Performance Management. Visit us today! http://pubads.g.doubleclick.net/gampad/clk?id=48897511&iu=/4140/ostg.clktrk _______________________________________________ Scikit-learn-general mailing list [email protected] https://lists.sourceforge.net/lists/listinfo/scikit-learn-general
