[dask] Support more meta data. #6132
Conversation
trivialfis
commented
Sep 18, 2020
trivialfis
commented
- label_lower_bound
- label_upper_bound
- base_margin
* Add base_margin, label_(lower|upper)_bound.
|
@hcho3 I modified the test for survival training to use it for dask. Could you please help taking a look? |
Codecov Report
@@ Coverage Diff @@
## master #6132 +/- ##
==========================================
+ Coverage 78.67% 78.83% +0.16%
==========================================
Files 12 12
Lines 3071 3090 +19
==========================================
+ Hits 2416 2436 +20
+ Misses 655 654 -1
Continue to review full report at Codecov.
|
| assert shap.ndim == 2 | ||
| assert shap.shape[0] == kRows | ||
| assert shap.shape[1] == kCols + 1 | ||
|
|
||
|
|
||
| def test_aft_survival(): | ||
| # survival doesn't handle empty dataset well. |
There was a problem hiding this comment.
I did not know this. How can we fix this?
There was a problem hiding this comment.
@hcho3 There are checks:
CHECK_NE(info.labels_lower_bound_.Size(), 0U)
<< "labels_lower_bound cannot be empty";
CHECK_NE(info.labels_upper_bound_.Size(), 0U)
<< "labels_upper_bound cannot be empty";in survival_metric.cu::EvalEWiseSurvivalBase::Eval.
There was a problem hiding this comment.
Should we remove the checks and have the objective return zero?
There was a problem hiding this comment.
Depends on the synchronization I think.
There was a problem hiding this comment.
AFT is an element-wise objective, so it should behave like reg:squarederror when it comes to handling empty datasets. Let me create a new issue for this.
