Fea hungarian expose precision #1673
Conversation
ChuckHastings
added
3 - Ready for Review
improvement
ChuckHastings
changed the title
|
rerun tests |
Codecov Report
@@ Coverage Diff @@
## branch-21.08 #1673 +/- ##
==============================================
Coverage ? 0.22%
==============================================
Files ? 80
Lines ? 3542
Branches ? 0
==============================================
Hits ? 8
Misses ? 3534
Partials ? 0 Continue to review full report at Codecov.
|
|
rerun tests |
| vertex_t num_rows, | ||
| vertex_t num_columns, | ||
| vertex_t *assignment, | ||
| weight_t precision); |
There was a problem hiding this comment.
IIRC, we used epsilon in raft and we're using precision here, I got they are two different functions, but both epsilon and precision are used to evaluate equality.... so better be consistent?
| weight_t const *costs, | ||
| vertex_t num_rows, | ||
| vertex_t num_columns, | ||
| vertex_t *assignment, |
There was a problem hiding this comment.
assignment"s"? (as this is an array of values and we're using cost"s", without "s", one may think cost"s" is an array and assignment is a single value).
| weight_t{0}, | ||
| thrust::maximum<weight_t>()); | ||
|
|
||
| rmm::device_uvector<weight_t> tmp_cost(n * n, handle.get_stream_view()); |
There was a problem hiding this comment.
Why use _v for tmp_row_assignment_v and tmp_col_assignment_v but no _v for tmp_cost? IIRC, you use _v for vectors and better be consistent at least within a code block.
|
@gpucibot merge |
Completes the work of exposing the precision parameter to the caller in python as requested by #1645 Fixes a few PR comments from PR #1673 that were deferred as that PR required merging to fix the build. Authors: - Chuck Hastings (https://github.com/ChuckHastings) Approvers: - Seunghwa Kang (https://github.com/seunghwak) - Brad Rees (https://github.com/BradReesWork) URL: #1674
Closes #1645
Closes #1646
Expose the precision parameter (epsilon in the Date/Nagi implementation) of the Hungarian algorithm to be controllable by the user. Add support for rectangular matrices.
Will be enabled for CI after rapidsai/raft#275 is merged.