Open adamgayoso opened 1 year ago
I think it would make sense to compare metric reimplementations to the original metrics and then make a call on picking one set in successive iterations of open problems.
We are testing for metric value equivalence in our testing suite
Okay. I guess this might still change when testing across datasets. I think @mumichae discovered some differences on our end... mainly for the clustering-based methods that now use k-means.
Yes we added a kmeans metric using known K, but are still working on speed
We also implemented the scib way with optimized leiden (instead of louvain) clustering
this one should be significantly faster than scib's implementation because:
For the batch integration embedding task we have developed this package:
https://scib-metrics.readthedocs.io/en/stable/
with python-only, jax-based implementations. This would easily allow fast clisi/ilisi. We also sped up:
https://scib-metrics.readthedocs.io/en/stable/generated/scib_metrics.nmi_ari_cluster_labels_leiden.html#scib_metrics.nmi_ari_cluster_labels_leiden
by using joblib and a faster leiden implementation in igraph.