Closed ogrisel closed 10 years ago
The l1_ratio
semantic change might have broken the mrec_tune example (at least on the Movie Lens 10M data). I need to investigate.
Thanks. Do you have any parameters for the l1 and l2 parameter that used to work well on the ML-10M dataset with sklearn <= 0.14.1 along with the cross-validated MRR so that I check that I can reproduce the result with the future sklearn 0.15?
In particular the meaning of the
l1_ratio
parameter is wrong in 0.14.1 and is going to be fixed in the new release.I also made it possible to train a model sequentially with
num_engines == 0
. This is useful for debugging with ipdb for instance.