Closed CyanideCentral closed 4 months ago
I have been able to create a graph that reproduces this error. I will investigate.
I have isolated the failing call. Working on a bug fix.
https://github.com/rapidsai/raft/pull/2319 should address this. Once it gets deployed we will retest internally. Goal is to address this in 24.06 release.
I have verified in the 24.06 branch that this bug fix eliminates the crash.
So this should be corrected in 24.06. If you are testing with conda packages, you can pull the latest nightly builds. If you are testing from source you should be able to do a clean build from source to pick up this change. If you are testing from released versions, version 24.06 will be released next month.
I have verified in the 24.06 branch that this bug fix eliminates the crash.
So this should be corrected in 24.06. If you are testing with conda packages, you can pull the latest nightly builds. If you are testing from source you should be able to do a clean build from source to pick up this change. If you are testing from released versions, version 24.06 will be released next month.
Thanks for your effort! We have tested 24.06 and can confirm that this issue is fixed.
Version
24.04
Which installation method(s) does this occur on?
Pip
Describe the bug.
I am using spectralModularityMaximizationClustering on a graph with 2.4M nodes and 110M nonzero entries in its adjacency matrix. In the following code, the intended number of clusters is k=47.
I started with num_eigen_vects=47=k, but encountered the following error. (Is this due to insufficient buffer size configured for the eigenvectors?)
Referring to Issue #710, I set smaller values for num_eigen_vects, and the algorithm only finishes without crashing when nev<=4. Still, with nev=4, the result contains 36 clusters instead of k=47.
In addition, I experimented with a subgraph consisting of the first 100k nodes, and the algorithm can finish with nev=47. For the subgraph with the first 1M nodes, the algorithm can only finish with nev<=16.
p.s. I first tried to instantiate the Graph object with adjlist from the CSR sparse matrix in the following code. However, directly calling spectralModularityMaximizationClustering afterward failed with the error
'NoneType' object has no attribute 'edgelist_df'
.Minimum reproducible example
No response
Relevant log output
No response
Environment details
Other/Misc.
No response
Code of Conduct