Ctrl+K

Attention

The vector search and clustering algorithms in RAFT are being migrated to a new library dedicated to vector search called cuVS. We will continue to support the vector search algorithms in RAFT during this move, but will no longer update them after the RAPIDS 24.06 (June) release. We plan to complete the migration by RAPIDS 24.08 (August) release.

Neighborhood Model Scoring#

Trustworthiness#

#include <raft/stats/trustworthiness.cuh>

namespace raft::stats

template<raft::distance::DistanceType distance_type, typename value_t, typename idx_t>
double trustworthiness_score(raft::resources const &handle, raft::device_matrix_view<const value_t, idx_t, raft::row_major> X, raft::device_matrix_view<const value_t, idx_t, raft::row_major> X_embedded, int n_neighbors, int batch_size = 512)#

Compute the trustworthiness score.

Note

The constness of the data in X_embedded is currently casted away and the data is slightly modified.

Template Parameters:

  • value_t – the data type

  • idx_t – Integer type used to for addressing

Parameters:

  • handle – [in] the raft handle

  • X – [in] Data in original dimension

  • X_embedded – [in] Data in target dimension (embedding)

  • n_neighbors – [in] Number of neighbors considered by trustworthiness score

  • batch_size – [in] Batch size

Returns:

Trustworthiness score

Neighborhood Recall#

#include <raft/stats/neighborhood_recall.cuh>

namespace raft::stats

template<typename IndicesValueType, typename IndexType, typename ScalarType, typename DistanceValueType = float>
void neighborhood_recall(raft::resources const &res, raft::device_matrix_view<const IndicesValueType, IndexType, raft::row_major> indices, raft::device_matrix_view<const IndicesValueType, IndexType, raft::row_major> ref_indices, raft::device_scalar_view<ScalarType> recall_score, std::optional<raft::device_matrix_view<const DistanceValueType, IndexType, raft::row_major>> distances = std::nullopt, std::optional<raft::device_matrix_view<const DistanceValueType, IndexType, raft::row_major>> ref_distances = std::nullopt, std::optional<raft::host_scalar_view<const DistanceValueType>> eps = std::nullopt)#

Calculate Neighborhood Recall score on the device for indices, distances computed by any Nearest Neighbors Algorithm against reference indices, distances. Recall score is calculated by comparing the total number of matching indices and dividing that value by the total size of the indices matrix of dimensions (D, k). If distance matrices are provided, then non-matching indices could be considered a match if abs(dist, ref_dist) < eps.

Usage example: 

raft::device_resources res;
// assume D rows and N column dataset
auto k = 64;
auto indices = raft::make_device_matrix<int>(res, D, k);
auto distances = raft::make_device_matrix<float>(res, D, k);
// run ANN algorithm of choice

auto ref_indices = raft::make_device_matrix<int>(res, D, k);
auto ref_distances = raft::make_device_matrix<float>(res, D, k);
// run brute-force KNN for reference

auto scalar = 0.0f;
auto recall_score = raft::make_device_scalar(res, scalar);

raft::stats::neighborhood_recall(res,
                                 raft::make_const_mdspan(indices.view()),
                                 raft::make_const_mdspan(ref_indices.view()),
                                 recall_score.view(),
                                 raft::make_const_mdspan(distances.view()),
                                 raft::make_const_mdspan(ref_distances.view()));

Template Parameters:

  • IndicesValueType – data-type of the indices

  • IndexType – data-type to index all matrices

  • ScalarType – data-type to store recall score

  • DistanceValueType – data-type of the distances

Parameters:
template<typename IndicesValueType, typename IndexType, typename ScalarType, typename DistanceValueType = float>
void neighborhood_recall(raft::resources const &res, raft::device_matrix_view<const IndicesValueType, IndexType, raft::row_major> indices, raft::device_matrix_view<const IndicesValueType, IndexType, raft::row_major> ref_indices, raft::host_scalar_view<ScalarType> recall_score, std::optional<raft::device_matrix_view<const DistanceValueType, IndexType, raft::row_major>> distances = std::nullopt, std::optional<raft::device_matrix_view<const DistanceValueType, IndexType, raft::row_major>> ref_distances = std::nullopt, std::optional<raft::host_scalar_view<const DistanceValueType>> eps = std::nullopt)#

Calculate Neighborhood Recall score on the host for indices, distances computed by any Nearest Neighbors Algorithm against reference indices, distances. Recall score is calculated by comparing the total number of matching indices and dividing that value by the total size of the indices matrix of dimensions (D, k). If distance matrices are provided, then non-matching indices could be considered a match if abs(dist, ref_dist) < eps.

Usage example: 

raft::device_resources res;
// assume D rows and N column dataset
auto k = 64;
auto indices = raft::make_device_matrix<int>(res, D, k);
auto distances = raft::make_device_matrix<float>(res, D, k);
// run ANN algorithm of choice

auto ref_indices = raft::make_device_matrix<int>(res, D, k);
auto ref_distances = raft::make_device_matrix<float>(res, D, k);
// run brute-force KNN for reference

auto scalar = 0.0f;
auto recall_score = raft::make_host_scalar(scalar);

raft::stats::neighborhood_recall(res,
                                 raft::make_const_mdspan(indices.view()),
                                 raft::make_const_mdspan(ref_indices.view()),
                                 recall_score.view(),
                                 raft::make_const_mdspan(distances.view()),
                                 raft::make_const_mdspan(ref_distances.view()));

Template Parameters:

  • IndicesValueType – data-type of the indices

  • IndexType – data-type to index all matrices

  • ScalarType – data-type to store recall score

  • DistanceValueType – data-type of the distances

Parameters:
On this page
Show Source