LoRANN Experiments

Experiments for the paper

Jääsaari, E., Hyvönen, V., & Roos, T. (2024). LoRANN: Low-Rank Matrix Factorization for Approximate Nearest Neighbor Search. Advances in Neural Information Processing Systems, 37.

The code implementing LoRANN is in a separate repository.

This project is a fork of the ANN-benchmarks project with new data sets and support for GPU experiments.

---

Requirements:

• Python 3.10 or newer
• Docker
• For GPU experiments, an NVIDIA GPU is required

Installation:

python3 -m pip install -r requirements.txt

Usage:

To build an algorithm, run e.g. python3 install.py --algorithm lorann.

To build all algorithms, run

for algo in faiss faiss_gpu glass hnswlib lorann lorann_gpu mrpt pynndescent qsg_ngt raft scann; do
  python3 install.py --algorithm $algo
done

To run an algorithm for e.g. the data set fashion-mnist-784-euclidean, run

python3 run.py --dataset fashion-mnist-784-euclidean --algorithm lorann --count 100 --parallelism 6

To plot results for the data set, run

python3 plot.py --dataset fashion-mnist-784-euclidean --count 100 --y-scale log

For a list of all the data sets, refer to the end of the file ann_benchmarks/datasets.py.

Our main experiments are performed on AWS r6i.4xlarge instances using Intel Xeon 8375C (Ice Lake) processors with hyperthreading disabled. To run our GPU experiments, we use an AWS g5.2xlarge instance with an NVIDIA A10G GPU (24 GB VRAM) and a mac2-m2pro.metal instance with an Apple M2 Pro SoC.