Skip to main content

Scalable Approximate Nearest Neighbor search library

Project description

ScaNN

ScaNN (Scalable Nearest Neighbors) is a method for efficient vector similarity search at scale. This code release implements [1], which includes search space pruning and quantization for Maximum Inner Product Search and also supports other distance functions such as Euclidean distance. The implementation is designed for x86 processors with AVX2 support. ScaNN achieves state-of-the-art performance on ann-benchmarks.com as shown on the glove-100-angular dataset below:

glove-100-angular

ScaNN can be configured to fit datasets with different sizes and distributions. It has both TensorFlow and Python APIs. The library shows strong performance with large datasets [1]. The code is released for research purposes. For more details on the academic description of algorithms, please see [1].

Reference [1]:

@inproceedings{avq_2020,
  title={Accelerating Large-Scale Inference with Anisotropic Vector Quantization},
  author={Guo, Ruiqi and Sun, Philip and Lindgren, Erik and Geng, Quan and Simcha, David and Chern, Felix and Kumar, Sanjiv},
  booktitle={International Conference on Machine Learning},
  year={2020},
  URL={https://arxiv.org/abs/1908.10396}
}

Installation

manylinux2014-compatible wheels are available on PyPI:

pip install scann

ScaNN supports Linux environments running Python versions 3.7-3.9. See docs/releases.md for release notes; the page also contains download links for ScaNN wheels prior to version 1.1.0, which were not released on PyPI.

In accordance with the manylinux2014 specification, ScaNN requires libstdc++ version 3.4.19 or above from the operating system. See here for an example of how to find your system's libstdc++ version; it can generally be upgraded by installing a newer version of g++.

Integration with TensorFlow Serving

We provide custom Docker images of TF Serving that are linked to the ScaNN TF ops. See the tf_serving directory for further information.

Building from source

To build ScaNN from source, first install the build tool bazel, Clang 8, and libstdc++ headers for C++17 (which are provided with GCC 9). Additionally, ScaNN requires a modern version of Python (3.7.x or later) and Tensorflow 2.7 installed on that version of Python. Once these prerequisites are satisfied, run the following command in the root directory of the repository:

python configure.py
CC=clang-8 bazel build -c opt --features=thin_lto --copt=-mavx2 --copt=-mfma --cxxopt="-D_GLIBCXX_USE_CXX11_ABI=0" --cxxopt="-std=c++17" --copt=-fsized-deallocation --copt=-w :build_pip_pkg
./bazel-bin/build_pip_pkg

A .whl file should appear in the root of the repository upon successful completion of these commands. This .whl can be installed via pip.

Usage

See the example in docs/example.ipynb. For a more in-depth explanation of ScaNN techniques, see docs/algorithms.md.

Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Source Distributions

No source distribution files available for this release.See tutorial on generating distribution archives.

Built Distributions

scann-1.2.4-cp39-cp39-manylinux2014_x86_64.whl (10.6 MB view hashes)

Uploaded CPython 3.9

scann-1.2.4-cp38-cp38-manylinux2014_x86_64.whl (10.6 MB view hashes)

Uploaded CPython 3.8

scann-1.2.4-cp37-cp37m-manylinux2014_x86_64.whl (10.6 MB view hashes)

Uploaded CPython 3.7m

Supported by

AWS AWS Cloud computing and Security Sponsor Datadog Datadog Monitoring Fastly Fastly CDN Google Google Download Analytics Microsoft Microsoft PSF Sponsor Pingdom Pingdom Monitoring Sentry Sentry Error logging StatusPage StatusPage Status page