All Neighbors

Source header: cuvs/neighbors/all_neighbors.hpp

The all-neighbors algorithm parameters.

GraphBuildParams

The all-neighbors algorithm parameters.

1
using GraphBuildParams = std::variant<graph_build_params::ivf_pq_params,
2
graph_build_params::nn_descent_params,
3
graph_build_params::brute_force_params>;

neighbors::all_neighbors::all_neighbors_params

Parameters used to build an all-neighbors graph (find nearest neighbors for all the

training vectors). For scalability, the all-neighbors graph construction algorithm partitions a set of training vectors into overlapping clusters, computes a local knn graph on each cluster, and merges the local graphs into a single global graph. Device memory usage and accuracy can be configured by changing the overlap_factor and n_clusters. The algorithm used to build each local graph is also configurable.

1
struct all_neighbors_params { ... };

Fields

The all-neighbors knn graph build

neighbors::all_neighbors::build

Builds an approximate all-neighbors knn graph (find nearest neighbors for all the

1
void build(
2
const raft::resources& handle,
3
const all_neighbors_params& params,
4
raft::host_matrix_view<const float, int64_t, row_major> dataset,
5
raft::device_matrix_view<int64_t, int64_t, row_major> indices,
6
std::optional<raft::device_matrix_view<float, int64_t, row_major>> distances      = std::nullopt,
7
std::optional<raft::device_vector_view<float, int64_t, row_major>> core_distances = std::nullopt,
8
float alpha                                                                       = 1.0);

training vectors)

Usage example:

compute core_distances. If core_distances is given, the resulting indices and distances will be mutual reachability space.

Parameters

Returns

void

Additional overload: neighbors::all_neighbors::build

Builds an approximate all-neighbors knn graph (find nearest neighbors for all the training

1
void build(
2
const raft::resources& handle,
3
const all_neighbors_params& params,
4
raft::device_matrix_view<const float, int64_t, row_major> dataset,
5
raft::device_matrix_view<int64_t, int64_t, row_major> indices,
6
std::optional<raft::device_matrix_view<float, int64_t, row_major>> distances      = std::nullopt,
7
std::optional<raft::device_vector_view<float, int64_t, row_major>> core_distances = std::nullopt,
8
float alpha                                                                       = 1.0);

vectors) params.n_clusters should be 1 for data on device. To use a larger params.n_clusters for efficient device memory usage, put data on host RAM.

Usage example:

compute core_distances. If core_distances is given, the resulting indices and distances will be mutual reachability space.

Parameters

Returns

void