[MRG+1] use class_weight through sample_weight in LogisticRegression with liblinear #6817
Conversation
| "solver cannot handle multiclass with " | ||
| "class_weight of type dict. Use the lbfgs, " | ||
| "newton-cg or sag solvers or set " | ||
| "class_weight='balanced'") |
There was a problem hiding this comment.
No, since liblinear solver handles sample_weight (#5274), we can make the behavior identical for all LogisticRegression's solvers, by merging the class_weight into the sample_weight.
|
+1 for merge |
| for class_weight in ({0: 0.1, 1: 0.2, 2: 0.5}, 'balanced'): | ||
| for n_classes in [3, 2]: | ||
| if n_classes == 2 and class_weight != 'balanced': | ||
| class_weight.pop(2) |
There was a problem hiding this comment.
to avoid bugs if this were extended, can we please do class_weight = class_weight.copy() first?
There was a problem hiding this comment.
I have changed the loops to avoid modifications on class_weight.
There was a problem hiding this comment.
I have changed the loops to avoid modifications on class_weight.
Yes, of course that's the sensible way to do it!
|
Otherwise, LGTM |
2c33e39 to
c640fce
Compare
|
would you like to add to |
c640fce to
1107f22
Compare
|
Done. I also added |
|
Let's do it. |
I forgot to update #5008 after we merged #5274 (cf. #5008 (comment)).
It makes the
class_weightbehavior consistent with all solver ofLogisticRegression, by putting theclass_weightinto thesample_weight.Fixes #6795