Reductive
Training of optimized product quantizers
Training of optimized product quantizers requires a LAPACK
implementation. For this reason, training of the OPQ
and
GaussianOPQ
quantizers is feature-gated by the opq-train
feature.
opq-train
is automatically enabled by selecting a BLAS/LAPACK
implementation. The supported implementations are:
- OpenBLAS (feature:
openblas
) - Netlib (feature:
netlib
) - Intel MKL (feature:
intel-mk;
)
A backend can be selected as follows:
[]
= { = "0.3", = ["openblas"] }
Running tests
To run all tests, specify the BLAS/LAPACK implementation:
$ cargo test --verbose --features "openblas"
Multi-threaded OpenBLAS
reductive
uses Rayon to parallelize quantizer training. However,
multi-threaded OpenBLAS is known to
conflict
with application threading. Is you use OpenBLAS, ensure that threading
is disabled, for instance by setting the number of threads to 1:
$ export OPENBLAS_NUM_THREADS=1