scverse/rapids_singlecell

Issues with Leidenclustering

Closed this issue · 4 comments

There is an issue with Leiden Clustering in cugraph version 23.06.

The issue has been reported rapidsai/cugraph#3749.

For the meantime please use rapids-23.04

Leiden clustering returned to normal after I created a new rapids-23.04 environment.

It greatly speeds up leiden clustering on large datasets, thanks for the great work!

Rapids-23.08 still has this bug. If you need Leiden Clustering stay with Rapids-23.04

I looks like leiden clustering will be fixed with rapids-23.12

Just wanted to document another example of this leiden over-clustering. I used conda env create -f conda/rsc_rapids_24.02.yml which was in the installation instructions here. The following is comparing the results for ~300,000 cells.

Using the following rapids function:

rsc.tl.louvain(adata, resolution=0.6,key_added="louvain_0.6")
rsc.tl.louvain(adata, resolution=0.8,key_added="louvain_0.8")

Or the default:

sc.tl.louvain(adata, resolution=0.6,key_added="louvain_0.6_default")
sc.tl.louvain(adata, resolution=0.8,key_added="louvain_0.8_default")
Screenshot 2024-05-17 at 11 58 25 PM

I checked the above information and saw cugraph 24.02 and rapids 24.02. Looks like I will be trying with env create -f conda/rsc_rapids_24.04.yml tomorrow instead haha! So far the rest is working great