sklearn fit silently drops or errors when length of eval_set and sample_weights do not match

[mmccarty]

XGBModel.fit packages the evals into a list of DMatrix objects by iterating over a range generated by the length of the given eval_set. However, when the user provides eval_set=[(x1, y2), (x2, y2)], sample_weight_eval_set=[weight1] the evals list will have silently dropped the second set. When the user provides eval_set=[(x1, y2)], sample_weight_eval_set=[weight1, weight2] an index exception is raised.

We discovered this when discussing and testing a PR adding early_stopping to dask-xgboost. We are wondering what is the right thing to do in this case? Raise an exception? Or do nothing?

[trivialfis]

@mmccarty Ah I see. Will fix it. Thanks for the issue. ;-)

[trivialfis]

Just by the way, we adopted that dask interface into xgboost itself recently.

[mmccarty]

@trivialfis Thank you! Looking forward to trying out native dask.

[trivialfis]

Yup. Still some pending PRs for bug squashing since it's relatively new. But stay tuned.

[trivialfis]

See #5152 I want to remove the validation data set parameter and replace it with call back functions.