2012/3/6 Alexandre Gramfort <[email protected]>: >> a) sparse coding is about 2 orders of magnitude slower than competing >> implementations right now, making it kind of useless except in toy >> 1996-sized situations (I'm supposed to find a way to benchmark >> this for Alex, but I can tell you that the situation is fairly bad >> currently, compared to e.g. SPAMS; Olivier said this had something to >> do with a badly optimized convergence check) > > we'll never be able to reach the same level of performance unless you > move everything to Cython including the LARS. Also to be fair the > comparison with SPAMS should be done with the same number of CPUs. > SPAMS makes a massive use of OpenMP e.g. all the LARS are computed in > parallel in the sparse coding step. But if it's the case, I agree 2 > orders of magnitude is not acceptable.
I think was DWF calls sparse coding is the LASSO implemented with coordinate descent (sparse coding with a fixed dictionary). Not the MiniBatchDictionaryLearning class that uses LARS and online block coordinate descent for the dictionary update. But David please feel to confirm. BTW, to address this you should start with a script that demonstrate the perf issue and then do some profiling on it. -- Olivier http://twitter.com/ogrisel - http://github.com/ogrisel ------------------------------------------------------------------------------ Virtualization & Cloud Management Using Capacity Planning Cloud computing makes use of virtualization - but cloud computing also focuses on allowing computing to be delivered as a service. http://www.accelacomm.com/jaw/sfnl/114/51521223/ _______________________________________________ Scikit-learn-general mailing list [email protected] https://lists.sourceforge.net/lists/listinfo/scikit-learn-general
