Improvements:
Added a class parameter for a pre-computed nearest neighbor graph matrix (i.e., the CSR array used by PARC). We use adaptive neighbor graphs, where k is not the same for all elements, so this functionality is essential for us to use PARC.
Fixes:
Within run_toobig_subPARC(), the line hnsw = self.make_knn_struct(too_big=True, big_cluster=X_data) wasn't working because self.make_knn_struct() returned None. It returns the hnsw index now instead.
Improvements: Added a class parameter for a pre-computed nearest neighbor graph matrix (i.e., the CSR array used by PARC). We use adaptive neighbor graphs, where
kis not the same for all elements, so this functionality is essential for us to use PARC.Fixes: Within
run_toobig_subPARC(), the linehnsw = self.make_knn_struct(too_big=True, big_cluster=X_data)wasn't working becauseself.make_knn_struct()returnedNone. It returns the hnsw index now instead.