Reductive

Training of optimized product quantizers

Training of optimized product quantizers requires a LAPACK implementation. For this reason, training of the OPQ and GaussianOPQ quantized is feature-gated by the opq-train feature. Without the the opq-train feature, you can train the PQ quantizer and use pre-trained quantizers.

If you use the opq-train feature, you also have to select a BLAS/LAPACK implementation. The supported implementations are:

• OpenBLAS (feature: openblas)
• Netlib (feature: netlib)
• Intel MKL (feature: intel-mk;)

There is a feature for macOS Accelerate (accelerate). However, Accelerate does not currently provide the necessary LAPACK routines. This feature is present in case Accelerate adds the necessary routines.

The opq-train feature and a backend can be enabled as follows:

[dependencies]
reductive = { version = "0.1", features = ["opq-train", "openblas"] }

Running tests

To run all tests, enable the opq-train feature and specify the BLAS/LAPACK implementation:

$ cargo test --verbose --features "opq-train 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