RAPIDS implementation of Scanpy rank_genes_groups appears incorrect about rapids-single-cell-examples

[avantikalal]

@teju85 We checked this and the problem still exists, despite the cuML bug being resolved. @cjnolet is looking into it.

[avantikalal]

Possibly relevant cuML issue: rapidsai/cuml#2478

[cjnolet]

. Also, I note that while the CPU output is sorted by score (i.e., the top 50 diff. genes have high scores, and are sorted in decreasing order), the GPU output seems to be unsorted, and some of the scores are very low.

Are you referring to the resulting scores (.e.g, adata.uns['rank_genes_groups']['scores'])? For me, the output for both CPU and GPU seem to be unsorted.

Correction: When using penalty='none', the major axis is indeed sorted in both the GPU and CPU notebooks.

While we wait for the release of the fix for rapidsai/cuml#2478, we have a couple options:

1. Turn off regularization by passing penalty='none' into the rank_genes_groups functions for both CPU and GPU.
2. Set the C hyper-parameter to the number of elements in X as recommended in rapidsai/cuml#2478.

[teju85]

hey folks,
just curious. Since the above cuML issue has been fixed, did any of you get a chance rerun the code afterwards? Are you still facing this issue?

[avantikalal]

@teju85 thanks for the reminder, we'll check and get back to you.

[avantikalal]

This issue should be resolved now: rapidsai/cuml#3645

Will test and close.