fixes and changes to the ranking metrics computed on cpu #5380
Conversation
- auc/aucpr ranking metric accelerated on cpu - fixes to the auc/aucpr metrics next, i'll create a pr for a gpu metric registry to decouple the cpu and gpu implementations
|
What are the fixes specifically? |
src/metric/rank_metric.cc
Outdated
| } | ||
|
|
||
| // we need pos > 0 && neg > 0 | ||
| if (total_pos <= 0.0 || total_neg <= 0.0) { |
There was a problem hiding this comment.
So the fix is trigger a fatal error when only one label type exists locally? e.g. in a distributed partition.
There was a problem hiding this comment.
Nvm looks like it just ignores the partition.
There was a problem hiding this comment.
if labels aren't diverse across all groups even in a non distributed environment.
Codecov Report
@@ Coverage Diff @@
## master #5380 +/- ##
=======================================
Coverage 84.07% 84.07%
=======================================
Files 11 11
Lines 2411 2411
=======================================
Hits 2027 2027
Misses 384 384 Continue to review full report at Codecov.
|
| const auto &h_labels = info.labels_.ConstHostVector(); | ||
| const auto &h_preds = preds.ConstHostVector(); | ||
|
|
||
| #pragma omp parallel reduction(+:sum_auc, auc_error) if (ngroups > 1) |
There was a problem hiding this comment.
I spent some time working on the deterministic algorithm recently: #5361 . I would like to know whether a component is deterministic or not.
| // Same thread can work on multiple groups one after another; hence, resize | ||
| // the predictions array based on the current group | ||
| rec.resize(gptr[group_id + 1] - gptr[group_id]); | ||
| #pragma omp parallel for schedule(static) reduction(+:total_pos, total_neg) \ |
There was a problem hiding this comment.
We should add a test for non-omp build. I will work on this.
it may help if you view the changes by ignoring the whitespace changes.
next, i'll create a pr for a gpu metric registry to decouple the cpu and gpu implementations
@RAMitchell @trivialfis @rongou - please review