Add ANN refinement method #1038
Conversation
tfeher
added
feature request
|
rerun tests |
1 similar comment
|
rerun tests |
|
rerun tests |
2 similar comments
|
rerun tests |
|
rerun tests |
There was a problem hiding this comment.
Overall, looks good and easy to read to me!
I have some doubts about trying to adapt the ivf-flat though. Since ivf-flat is wired to accept the interleaved data, we end up reading the vectors twice, which does not sound very efficient. Could a naive segmented distance kernel followed by a select_topk invocation be faster and even less code? Perhaps, we could explore this in a follow-up.
There was a problem hiding this comment.
Thanks Artem for the review, I have updated the PR!
Since ivf-flat is wired to accept the interleaved data, we end up reading the vectors twice, which does not sound very efficient.
That is right. Actually the index building is the time consuming part because it need to gather the neighbor vectors based on their indices, so we do not have a good access pattern. Streaming over the already gathered data much faster.
Could a naive segmented distance kernel followed by a select_topk invocation be faster and even less code?
Yes, we shall explore that in a follow up work.
| raft::neighbors::ivf_flat::index<data_t, idx_t> refinement_index( | ||
| handle, metric, n_queries, false, dim); | ||
|
|
||
| raft::spatial::knn::ivf_flat::detail::fill_refinement_index(handle, |
There was a problem hiding this comment.
Just a little note: I'm hoping to remove the raft::spatial::knn namespace altogether in 23.02 and officially move all the detail headers over to raft::neighbors. Hopefully we can get all the deprecated files removed as well.
|
@gpucibot merge |
This PR implements refinement for approximate nearest neighbor search.
Refinement is a post processing step for ANN search, it follows an ANN search that returned
k0neighbor candidates,and select
kout of these candidates. The selection by calculating exact distances from the original dataset.Refinement can increase accuracy. It is useful for ANN methods that quantize the dataset and therefore loose accuracy during distance calculation (e.g. IVF-PQ).