cuml
cuml copied to clipboard
DBSCAN utilize rbc eps_neighbors
This PR enables rbc eps-neighbor computation via raft. The resulting adjacency matrix is sparse and allows to skip the implicit conversion.
Notes:
- the 'algorithm'-parameter was added to the DBSCAN init signature to allow the user to choose (default is 'brute', 'rbc' is optional)
- the memory management is still very conservative, assuming a dense adjacency matrix and therefore selecting comparably small batches
- in case maximum row length of a batch is sufficiently small the CSR structure can be computed in a single pass
CC @tfeher
rerun tests
@tfeher , thanks for reviewing. I just added a small correction and re-triggered the CI.
/test
rerun tests
/merge