Andreas wrote:

> Sorry for being terse, I should be working.

Yes, me too. I think this is the last commit you'll see from me for a while.

> We really can not break backward compatibility.
> One possibility is to have the old ``grid_scores_`` be the same as

Backwards-compatible ``grid_scores_`` is available again, and tested.

> it was before and make ``cv_scores_`` have the new structure.

Currently my solution stores precisely three attributes:
* grid_results_ -- data on a per-grid-point basis (1d struct array)
* fold_results_ -- data on a per-fold basis (2d struct array)
* best_index_ -- the offset into those arrays for the highest-scoring data point
(also, best_estimator_ is stored if refit=True)

grid_scores_, best_params_ and best_score_ are now calculated as
properties since they are derivative from the above.
grid_ and fold_ may not be eminently clear, but "scores" is a misnomer
(because these include parameters and times).

Composite score output (at train and test time) is also now tested (19ea7eac0d).

I assume backwards-compatibility for fit_grid_point is not necessary.
However, it currently returns clf_params for no need.

Finally, if you don't like the Scorer.store() approach, it could be
replaced with Scorer.named_scores() which would return an iterable of
(name, value) pairs, one of those names necessarily being 'score'.
Then the prefixing happens on the grid_search side. In fact, I think
this is a bit neater...

https://github.com/jnothman/scikit-learn/tree/grid_search_more_info

------------------------------------------------------------------------------
Everyone hates slow websites. So do we.
Make your web apps faster with AppDynamics
Download AppDynamics Lite for free today:
http://p.sf.net/sfu/appdyn_d2d_mar
_______________________________________________
Scikit-learn-general mailing list
[email protected]
https://lists.sourceforge.net/lists/listinfo/scikit-learn-general

Reply via email to