Le 13 avril 2012 15:55, Conrad Lee <[email protected]> a écrit : > Thanks for the suggestion of how to change RFE.py to exploit warm_start. > Should I add this feature to rfe.py and make a pull request? Or is this > functionality too specialized? > >> As we said in the issue you opened recently, SGDRegressor doesn't monitor >> convergence therefore warm_start won't make things faster out of the box. >> But you can still try to reduce the number of iterations by hand (since the >> solution is warm-started, you can expect SGD to find a solution in a fewer >> iterations). > > > Do you know of other fast estimators I can use that do check for > convergence?
We should improve SGD models to have early stopping with or without explicit validation sets. Nobody is working on this AFAIK. > Or is SGDRegressor in general the fastest? I know speed will depend on many > factors, what other estimators would you suggest as quite quick? It depends on the data: SGDRegressor is good at n_samples >> n_features. Lasso or LassoLARS are good at n_features >> n_samples with very few informative features, RidgeRegression is also fast in some if n_features and n_samples are very different. -- Olivier http://twitter.com/ogrisel - http://github.com/ogrisel ------------------------------------------------------------------------------ For Developers, A Lot Can Happen In A Second. Boundary is the first to Know...and Tell You. Monitor Your Applications in Ultra-Fine Resolution. Try it FREE! http://p.sf.net/sfu/Boundary-d2dvs2 _______________________________________________ Scikit-learn-general mailing list [email protected] https://lists.sourceforge.net/lists/listinfo/scikit-learn-general
