[MRG] Adds support for sample weight in hinge loss #3788
Conversation
4c0b04b to
129733f
Compare
|
Can anyone please review this? cc @MechCoder @jnothman @arjoly |
| @@ -1184,6 +1184,32 @@ def test_hinge_loss_multiclass_invariance_lists(): | |||
| dummy_hinge_loss) | |||
|
|
|||
|
|
|||
| def test_hinge_loss_multiclass_sample_weight(): | |||
There was a problem hiding this comment.
Do you have reason to believe the invariance test is insufficient?
There was a problem hiding this comment.
You are right. Invariance test is sufficient here. I will remove this in next commit.
|
The implementation looks good to me, but I'm not sure the specialised test adds anything. |
| @@ -1492,4 +1495,7 @@ def hinge_loss(y_true, pred_decision, labels=None): | |||
| losses = 1 - margin | |||
| # The hinge_loss doesn't penalize good enough predictions. | |||
| losses[losses <= 0] = 0 | |||
| return np.mean(losses) | |||
| if sample_weight is None: | |||
There was a problem hiding this comment.
This should not be necessary. np.average already handles the case where weights=None
|
Referencing #3450 |
|
Need one more reviewer. @MechCoder Can you please have a look at this? |
|
This looks good, except for the fact that you should validate the first dimension of the pred_decision and sample_weights. This can be done just be adding |
|
+1 when addressing @MechCoder's comment. |
|
Merged as d21d6dd . Thanks. |
|
Thanks everyone for the reviews. :-) |
This is dependent on #3607. I will uncomment the added lines after it's merge in main tree.