Comments for I say things http://fseoane.net/blog mostly about programming, machine learning and such Sun, 06 Nov 2011 19:17:35 +0000 hourly 1 http://wordpress.org/?v=3.1.2 Comment on Low rank approximation by Mathieu http://fseoane.net/blog/2011/low-rank-approximation/#comment-75720 Mathieu Sun, 06 Nov 2011 19:17:35 +0000 http://fseoane.net/blog/?p=1128#comment-75720 Nice! Here's the compression rate for rank 50: >>> (600 * 50 + 50 + 50 * 400) * 1.0 / (600 * 400) 0.20854166666666665 Nice! Here’s the compression rate for rank 50:

>>> (600 * 50 + 50 + 50 * 400) * 1.0 / (600 * 400)
0.20854166666666665

]]> Comment on qr_multiply function in scipy.linalg by Fabian http://fseoane.net/blog/2011/qr_multiply-function-in-scipy-linalg/#comment-75474 Fabian Thu, 03 Nov 2011 11:04:02 +0000 http://fseoane.net/blog/?p=1062#comment-75474 Hey, The reason why it is so effective is because when LAPACK computes the QR factorization of a matrix it doesn't explicitly compute the Q matrix but stores it a compressed format they call ``elementary reflectors``. This function performs multiplication directly on those elementary reflectors, thus avoiding to compute the Q matrix explicitly. I don't think they implement the machinery for sparse matrices, but sounds feasible to implement a dense-sparse qr_multiply(A, v) function, where A is dense but v is sparse. Hey,

The reason why it is so effective is because when LAPACK computes the QR factorization of a matrix it doesn’t explicitly compute the Q matrix but stores it a compressed format they call “elementary reflectors“.

This function performs multiplication directly on those elementary reflectors, thus avoiding to compute the Q matrix explicitly.

I don’t think they implement the machinery for sparse matrices, but sounds feasible to implement a dense-sparse qr_multiply(A, v) function, where A is dense but v is sparse.

]]> Comment on qr_multiply function in scipy.linalg by Mathieu http://fseoane.net/blog/2011/qr_multiply-function-in-scipy-linalg/#comment-75455 Mathieu Thu, 03 Nov 2011 06:34:17 +0000 http://fseoane.net/blog/?p=1062#comment-75455 That sounds very useful. Can you give more details how does it work? Would that work with sparse matrices too? That sounds very useful. Can you give more details how does it work? Would that work with sparse matrices too?

]]> Comment on django, change language settings dynamically by green witch http://fseoane.net/blog/2009/django-change-language-settings-dynamically/#comment-74826 green witch Tue, 25 Oct 2011 12:12:35 +0000 http://fseoane.net/blog/?p=160#comment-74826 Thanks a lot bro.. Thanks a lot bro..

]]> Comment on Computing the vector norm by Adam http://fseoane.net/blog/2011/computing-the-vector-norm/#comment-73197 Adam Tue, 04 Oct 2011 02:28:23 +0000 http://fseoane.net/blog/?p=440#comment-73197 Thanks, this is a huge improvement. Why isn't the numpy norm implemented that way? Here's a test comparing the SCIPY BLAS, numpy, and "dot" routine speeds: <a href="http://tinypic.com?ref=4qg10n" rel="nofollow"></a> Thanks, this is a huge improvement. Why isn’t the numpy norm implemented that way?

Here’s a test comparing the SCIPY BLAS, numpy, and “dot” routine speeds:

]]> Comment on Reworked example gallery for scikit-learn by Ralf http://fseoane.net/blog/2011/reworked-example-gallery-for-scikit-learn/#comment-71218 Ralf Fri, 09 Sep 2011 05:34:28 +0000 http://fseoane.net/blog/?p=1017#comment-71218 That looks fantastic! I really need to find a reason to use scikit-learn more often:) That looks fantastic! I really need to find a reason to use scikit-learn more often:)

]]> Comment on A profiler for Python extensions by fabian http://fseoane.net/blog/2011/a-profiler-for-python-extensions/#comment-70829 fabian Mon, 05 Sep 2011 17:08:40 +0000 http://fseoane.net/blog/?p=531#comment-70829 That info is useful, thanks! That info is useful, thanks!

]]> Comment on A profiler for Python extensions by Tyberius Prime http://fseoane.net/blog/2011/a-profiler-for-python-extensions/#comment-70803 Tyberius Prime Mon, 05 Sep 2011 09:52:10 +0000 http://fseoane.net/blog/?p=531#comment-70803 Nevermind - doing 'google-pprof my_extension.so output.prof' instead of 'google-pprof /usr/bin/python output.prof' is what I should have been doing ;). Nevermind – doing ‘google-pprof my_extension.so output.prof’ instead of ‘google-pprof /usr/bin/python output.prof’ is what I should have been doing ;) .

]]> Comment on A profiler for Python extensions by Tyberius Prime http://fseoane.net/blog/2011/a-profiler-for-python-extensions/#comment-70797 Tyberius Prime Mon, 05 Sep 2011 09:23:56 +0000 http://fseoane.net/blog/?p=531#comment-70797 Thanks Fabian. I'm on Ubuntu lucid, 64 bit. google-pprof is pretty much what I did - though I have the same issue if I pass '-v' (which I guess only calls pprof?). My extension is in a weird subdirectory, but copying the .so to the current one did not help. Thanks Fabian. I’m on Ubuntu lucid, 64 bit.

google-pprof is pretty much what I did – though I have the same issue if I pass ‘-v’ (which I guess only calls pprof?).
My extension is in a weird subdirectory, but copying the .so to the current one did not help.

]]> Comment on A profiler for Python extensions by Fabian http://fseoane.net/blog/2011/a-profiler-for-python-extensions/#comment-70666 Fabian Sat, 03 Sep 2011 06:11:45 +0000 http://fseoane.net/blog/?p=531#comment-70666 Tyberius Prime: On what OS are you working ? I found that memory address is sometimes not resolved correctly under OSX ... You could also try to read the .prof file directly with pprof (or google-pprof depending on the system) Tyberius Prime: On what OS are you working ? I found that memory address is sometimes not resolved correctly under OSX …

You could also try to read the .prof file directly with pprof (or google-pprof depending on the system)

]]> Comment on A profiler for Python extensions by Tyberius Prime http://fseoane.net/blog/2011/a-profiler-for-python-extensions/#comment-70614 Tyberius Prime Fri, 02 Sep 2011 15:16:29 +0000 http://fseoane.net/blog/?p=531#comment-70614 Love the idea of the module, and the basics are working, but all I get in the output are addresses, not resolved function names (for my extension at least. System calls do get resolved). Is there anything special I have to keep in mind when compiling my extension? I am passing '-g' to gcc. passing --add_lib to pprof does not help either :( Thanks for you work! So long, Tyberius Prime Love the idea of the module,
and the basics are working, but all I get in the output are addresses, not resolved function names (for my extension at least. System calls do get resolved).

Is there anything special I have to keep in mind when compiling my extension? I am passing ‘-g’ to gcc.
passing –add_lib to pprof does not help either :(

Thanks for you work!
So long,
Tyberius Prime

]]> Comment on scikit-learn’s EuroScipy 2011 coding sprint — day two by Thouis Jones http://fseoane.net/blog/2011/scikit-learn%e2%80%99s-euroscipy-2011-coding-sprint-day-two/#comment-70276 Thouis Jones Mon, 29 Aug 2011 10:45:23 +0000 http://fseoane.net/blog/?p=999#comment-70276 Sad I missed the sprint, but I'm glad that the BallTree code got replaced. Wrapping the C++ was a good learning experience (I'm going to look back to that bit of history in the git repo whenever I have to do it again), but not something that one would want to keep around, if possible. Sad I missed the sprint, but I’m glad that the BallTree code got replaced. Wrapping the C++ was a good learning experience (I’m going to look back to that bit of history in the git repo whenever I have to do it again), but not something that one would want to keep around, if possible.

]]> Comment on solve triangular matrices using scipy.linalg by vene http://fseoane.net/blog/2010/solve_triangular-in-scipy-linalg/#comment-68905 vene Thu, 11 Aug 2011 14:31:19 +0000 http://fseoane.net/blog/?p=192#comment-68905 What is the Axiom? What is the Axiom?

]]> Comment on Ridge regression path by Ralph Leonhardt http://fseoane.net/blog/2011/ridge-regression-path/#comment-67889 Ralph Leonhardt Fri, 29 Jul 2011 06:54:41 +0000 http://fseoane.net/blog/?p=939#comment-67889 Ist there a chance to link this to the predictive performance of the classifier, for example the MSE? That would be interesting. Statistically speaking, Ridge Regressions live from the trade-off between bias and variance as compared to the best unbiased learners. Ist there a chance to link this to the predictive performance of the classifier, for example the MSE? That would be interesting. Statistically speaking, Ridge Regressions live from the trade-off between bias and variance as compared to the best unbiased learners.

]]> Comment on Ridge regression path by serch http://fseoane.net/blog/2011/ridge-regression-path/#comment-67010 serch Tue, 19 Jul 2011 14:25:46 +0000 http://fseoane.net/blog/?p=939#comment-67010 This is truly great Fabián, keep up the good work. This is truly great Fabián, keep up the good work.

]]> Comment on Ridge regression path by Mathieu http://fseoane.net/blog/2011/ridge-regression-path/#comment-66151 Mathieu Tue, 12 Jul 2011 15:12:59 +0000 http://fseoane.net/blog/?p=939#comment-66151 A solver="thin_svd" option would be a nice addition then. In RidgeCV, the trick could be used to do k-fold cross-validation. You would need to do only k SVD computations instead of k * |alphas|. A solver=”thin_svd” option would be a nice addition then. In RidgeCV, the trick could be used to do k-fold cross-validation. You would need to do only k SVD computations instead of k * |alphas|.

]]> Comment on Ridge regression path by fabian http://fseoane.net/blog/2011/ridge-regression-path/#comment-66129 fabian Tue, 12 Jul 2011 10:19:52 +0000 http://fseoane.net/blog/?p=939#comment-66129 Indeed, I will submit a patch for scikit-learn when I'm a bit more familiar with the code. Also, I found the current implementation (cholesky-based) to break down for (near) singular matrices and small parameters, while this SVD-based one is more robust, that's the main reason why I didn't use the scikit-learn implementation for these figures. Indeed, I will submit a patch for scikit-learn when I’m a bit more familiar with the code.

Also, I found the current implementation (cholesky-based) to break down for (near) singular matrices and small parameters, while this SVD-based one is more robust, that’s the main reason why I didn’t use the scikit-learn implementation for these figures.

]]> Comment on Ridge regression path by Mathieu http://fseoane.net/blog/2011/ridge-regression-path/#comment-66127 Mathieu Tue, 12 Jul 2011 10:02:03 +0000 http://fseoane.net/blog/?p=939#comment-66127 As always, nice-looking figures! Maybe we could add a ridge_path function to the ridge module to be consistent with lars and lasso? As always, nice-looking figures!

Maybe we could add a ridge_path function to the ridge module to be consistent with lars and lasso?

]]> Comment on LLE comes in different flavours by Andrea http://fseoane.net/blog/2011/lle-comes-in-different-flavours/#comment-65361 Andrea Sat, 02 Jul 2011 19:07:15 +0000 http://fseoane.net/blog/?p=882#comment-65361 it works great (I also updated my repo). Thank you very much to maintain this blog and I hope you keep on this way! it works great (I also updated my repo). Thank you very much to maintain this blog and I hope you keep on this way!

]]> Comment on LLE comes in different flavours by Ray Jones http://fseoane.net/blog/2011/lle-comes-in-different-flavours/#comment-65350 Ray Jones Sat, 02 Jul 2011 13:28:02 +0000 http://fseoane.net/blog/?p=882#comment-65350 Is it me, or are the ltsa and hessian outputs (almost?) exactly the same? Is it me, or are the ltsa and hessian outputs (almost?) exactly the same?

]]>