2012/6/28 Charles-Pierre Astolfi <[email protected]>: > Hi scikiters, > > I am trying to compute lasso/lars for several values of the > regularization parameter. > > A clean way to do this, is like this: >>>> alphas = [some long list of positive values] >>>> models = linear_model.lars_path(X, y, alphas=alphas)
Nope you cannot path the list of alphas to `lars_path` as it will automatically find the relevant inflection points analytically, see for instance this example : http://scikit-learn.org/stable/auto_examples/linear_model/plot_lasso_lars.html > Unfortunately, it seems to be very slow, compared to lars_path + > manual interpolation on each alpha. Ok so in the previous comment you probably meant lasso_path instead of lars_path. As in: http://scikit-learn.org/stable/auto_examples/linear_model/plot_lasso_coordinate_descent_path.html > My understanding is that it is > because the returned model is an ElasticNet and, as such, cannot use > lars regularization path to efficiently compute the solutions for > several alphas. Yes it is an iterated Coordinate Descent optimization of the LASSO objective function rather than a LASSO/LARS optimizer but it does warm restarts so it should still be competitive with LARS. However I think that the convergence checks in the current implementation are expensive right now so fine grids my still be expensive. There are two ongoing GSoC projects that might fix this perf issue at some point. > Am I missing something? You might now be aware that the Lasso CD optimization is very expensive when alpha is getting very small. Try to truncate your alphas list down to the smallest values that are still easy to optimize for. > If I need speed, should I interpolate myself? It's probably a good idea and could be a nice new feature to contribute to scikit-learn. > Is it worth a pull request? Sure :) -- Olivier http://twitter.com/ogrisel - http://github.com/ogrisel ------------------------------------------------------------------------------ Live Security Virtual Conference Exclusive live event will cover all the ways today's security and threat landscape has changed and how IT managers can respond. Discussions will include endpoint security, mobile security and the latest in malware threats. http://www.accelacomm.com/jaw/sfrnl04242012/114/50122263/ _______________________________________________ Scikit-learn-general mailing list [email protected] https://lists.sourceforge.net/lists/listinfo/scikit-learn-general
