leiden nmi ari #24
leiden nmi ari #24
Conversation
Codecov Report
Additional details and impacted files@@ Coverage Diff @@
## main #24 +/- ##
==========================================
- Coverage 93.75% 93.23% -0.52%
==========================================
Files 9 9
Lines 288 325 +37
==========================================
+ Hits 270 303 +33
- Misses 18 22 +4
Flags with carried forward coverage won't be shown. Click here to find out more.
|
|
|
||
| def test_nmi_ari_cluster_labels_leiden_parallel(): | ||
| X, labels = dummy_x_labels(return_symmetric_positive=True) | ||
| nmi, ari = scib_metrics.nmi_ari_cluster_labels_leiden(X, labels, optimize_resolution=True, n_jobs=2) |
There was a problem hiding this comment.
is it possible to check this against some leiden impl to make sure the clusters are the same? Is this reliable for diff random seeds?
There was a problem hiding this comment.
ah jk i see its leiden vs louvain
There was a problem hiding this comment.
leiden is faster. And the igraph implementation is even faster than the one scanpy uses
src/scib_metrics/_ari_nmi.py
Outdated
| ) | ||
| except ImportError: | ||
| logger.info("Using for loop over resolutions. pip install joblib for parallelization.") | ||
| out = [nmi_ari_cluster_labels_leiden(X, labels, False, r) for r in resolutions] |
There was a problem hiding this comment.
this recursive pattern feels like a recipe for bugs. would be much cleaner to just split out the else case into a helper then call that from both cases (optimize or not optimize)
Adds leiden nmi ari scores to more closely match scib (they use louvain). This does a search of 10 resolutions of leiden clustering to pick the optimal NMI (as in scIB, but it uses 20 res params)