[MRG] DOC FIX: multi-target linear model attribute shapes #1584
Conversation
| @@ -440,15 +440,16 @@ class Lars(LinearModel, RegressorMixin): | |||
|
|
|||
| Attributes | |||
| ---------- | |||
| `coef_path_` : array, shape = [n_features, n_alpha] | |||
| `coef_path_` : array, shape = [n_features, n_alpha] | list of \ | |||
| n_targets such arrays | |||
There was a problem hiding this comment.
more than 80c are allowed in docstrings AFAIK. I've never seen a \ before in a docstring
There was a problem hiding this comment.
Yeah pep8 doesn't mind that i think.. also because address links in docstrings are super long sometimes and you can't really split them, so docstrings they can be long AFAIK
|
thanks for doing this. No time for me before next week ... or later :( |
|
Thanks for working on this. I'll try to find time to review it soon. Hopefully @GaelVaroquaux can give some more qualified feedback on the weekend. |
|
I think this is related: #1615 |
|
Any news? |
|
Is anything holding back the merge of this PR? |
|
I guess its incompleteness based on the two concerns in my original post. I can adress the second one easily. |
|
If you fix that, +1 for merging. We can open a new issue for the second concern. |
|
My thoughts exactly. |
|
Ok, feel free to merge when you're done with (1). |
|
does ElasticNetCV behave correctly with multi target Y ? I am not sure we should support it since all alphas will be different and the computation is likely to be as fast as when run for each target independently. |
|
@agramfort: I'm not sure what correct behaviour is, here. The fact that it doesn't crash is due to inheritance. I fear that due to the loop over targets, each consecutive alpha would override the previous. I have not investigated this, though. Even if the alpha is the one that's best overall, I am still not sure whether this should be called correct behaviour, as opposed to a list of |
DOC fixes and standardization in least_angle docstrings More standardization in linear model docs
|
I squashed and now there's just one commit to bring into master (though it cannot be done with the green button apparently). However I'm waiting for some consistency discussion that I started on the list. |
|
Merged by rebase. |
I have two concerns.
ElasticNetCVand derivated classes only select one point in hyperparameter space. It's not documented, and honestly I don't know at the moment how this point is selected when y is multi-target. I would expect it to be the overall best one over all targets.least_anglebased classes do not document thealphas_,active_andcoef_path_parameters, and as they are they can't be easily documented. Even thelars_pathfunction misdocuments them as being shaped aftermax_featureswhich is not true in the case of early stopping or especially in the case of lasso where features can be removed.I might not have time to do this the following few days but if nobody else can spare the time and it's holding back the release please ping me to do it.