Configure an approximate KNN query algorithm

For the main path, pass "ivfflat" or "ivfpq" directly to the algo argument of cuda_ml_knn(); cuda.ml then lets the backend choose the index parameters. Use these constructors only when those parameters need to be set explicitly.

Usage

cuda_ml_knn_algo_ivfflat(nlist, nprobe)

cuda_ml_knn_algo_ivfpq(nlist, nprobe, m, n_bits)

Arguments

nlist

Number of cells to partition dataset into.

nprobe

At query time, the number of cells used for approximate nearest neighbor search.

m

Number of subquantizers.

n_bits

Bits allocated per subquantizer, from 4 to 8. The product of m and n_bits must be divisible by 8.

Value

A KNN algorithm specification to pass to the algo argument of cuda_ml_knn().

Details

Both algorithms partition the training data into nlist cells and search nprobe cells for each query. IVFFlat stores the original vectors and therefore needs only those two parameters. IVFPQ also compresses vectors using product quantization, so it additionally requires the number of subquantizers (m) and the bits allocated to each subquantizer (n_bits). The distinct constructors keep the required parameters for each algorithm explicit.

See also

cuda_ml_knn()