Neighborhood Graph and Tree for Indexing High-dimensional Data

Home / Installation / Command / License / Publications / About Us / 日本語

NGT provides commands and a library for performing high-speed approximate nearest neighbor searches against a large volume of data in high dimensional vector data space (several ten to several thousand dimensions).

News

• 08/10/2022 QBG (Quantized Blob Graph) and QG (renewed NGTQG) are now available. The command-line interface ngtq and ngtqg are now obsolete by replacing qbg. (v2.0.0)
• 02/04/2022 FP16 (half-precision floating point) is now available. (v1.14.0)
• 03/12/2021 The results for the quantized graph are added to this README.
• 01/15/2021 NGT v1.13.0 to provide the quantized graph (NGTQG) is released.
• 11/04/2019 NGT tutorial has been released.
• 06/26/2019 Jaccard distance is available. (v1.7.6)
• 06/10/2019 PyPI NGT package v1.7.5 is now available.
• 01/17/2019 Python NGT can be installed via pip from PyPI. (v1.5.1)
• 12/14/2018 NGTQ (NGT with Quantization) is now available. (v1.5.0)
• 08/08/2018 ONNG is now available. (v1.4.0)

Methods

This repository provides the following methods.

• NGT: Graph and tree-based method
• QG: Quantized graph-based method
• QBG: Quantized blob graph-based method

Note: Since QG and QBG require BLAS and LAPACK libraries, if you use only NGT (Graph and tree-based method) without the additional libraries like V1, you can disable QB and QBG with this option.

Installation

Build

Downloads

• Releases

On Linux without QG and QBG

  $ unzip NGT-x.x.x.zip
  $ cd NGT-x.x.x
  $ mkdir build
  $ cd build
  $ cmake -DNGT_QBG_DISABLED=ON ..
  $ make
  $ make install
  $ ldconfig /usr/local/lib

On CentOS

  $ yum install blas-devel lapack-devel
  $ unzip NGT-x.x.x.zip
  $ cd NGT-x.x.x
  $ mkdir build
  $ cd build
  $ cmake ..
  $ make
  $ make install
  $ ldconfig /usr/local/lib

On Ubuntu

  $ apt install libblas-dev liblapack-dev
  $ unzip NGT-x.x.x.zip
  $ cd NGT-x.x.x
  $ mkdir build
  $ cd build
  $ cmake ..
  $ make
  $ make install
  $ ldconfig /usr/local/lib

On macOS using homebrew

  $ /usr/bin/ruby -e "$(curl -fsSL https://raw.githubusercontent.com/Homebrew/install/master/install)"
  $ brew install cmake
  $ brew install libomp
  $ unzip NGT-x.x.x.zip
  $ cd NGT-x.x.x
  $ mkdir build
  $ cd build
  $ cmake ..
  $ make
  $ make install

Pre-Built

On macOS

  $ brew install ngt

NGT (Graph and tree-based method)

Key Features

• Supported operating systems: Linux and macOS
• Object additional registration and removal are available.
• Objects beyond the memory size can be handled using the shared memory (memory mapped file) option.
• Supported distance functions: L1, L2, Cosine similarity, Angular, Hamming, Jaccard, Poincare, and Lorentz
• Data Types: 4 byte floating point number and 1 byte unsigned integer
• Supported languages: Python, Ruby, PHP, Rust, Go, C, and C++
• Distributed servers: ngtd and vald

Documents

• NGT tutorial

Utilities

• Command : ngt and qbg
• Server : ngtd and vald

Supported Programming Languages

• Python
• Ruby (Thanks Andrew!)
• PHP (Thanks Andrew!)
• Rust (Thanks Romain!)
• JavaScript/NodeJS : ngt-tool and spatial-db-ngt (Thanks stonkpunk!)
• Go
• C
• C++(sample code)

The following build parameters are available

Build parameters

Shared memory use

The index can be placed in shared memory with memory mapped files. Using shared memory can reduce the amount of memory needed when multiple processes are using the same index. In addition, it can not only handle an index with a large number of objects that cannot be loaded into memory, but also reduce time to open it. Since changes become necessary at build time, please add the following parameter when executing "cmake" in order to use shared memory.

  $ cmake -DNGT_SHARED_MEMORY_ALLOCATOR=ON ..

Note: Since there is no lock function, the index should be used only for reference when multiple processes are using the same index.

Large-scale data use

When you insert more than about 5 million objects for the graph-based method, please add the following parameter to improve the search time.

  $ cmake -DNGT_LARGE_DATASET=ON ..

Disable QG and QBG

QG and QBG require BLAS and LAPACK libraries. If you would not like to install these libraries and do not use QG and QBG, you can disable QG and QBG.

  $ cmake -DNGT_QBG_DISABLED=ON ..

QG (Quantized graph-based method)

Key Features

• Higher performance than the graph and tree-based method
• Supported operating systems: Linux and macOS
• Supported distance functions: L2 and Cosine similarity

Documents

• QG tutorial

Utilities

• Command : qbg

Supported Programming Languages

• C++
• C
• Python only for search

Build parameters

For QG, it is recommended to disable rotation of the vector space and residual vectors to improve performance as follows.

  $ cmake -DNGTQG_NO_ROTATION=ON -DNGTQG_ZERO_GLOBAL=ON ..

QBG (Quantized blob graph-based method)

Key Features

• QBG can handle billions of objects.
• Supported operating systems: Linux and macOS
• Supported distance functions: L2

Utilities

• Command : qbg

Supported Programming Languages

• C++
• C
• Python only for search

Benchmark Results

The followings are the results of ann benchmarks for NGT v2.0.0 where the timeout is 5 hours on an AWS c5.4xlarge instance.

glove-100-angular

gist-960-euclidean

fashion-mnist-784-euclidean

nytimes-256-angular

sift-128-euclidean

License

Copyright (C) 2015 Yahoo Japan Corporation

Licensed under the Apache License, Version 2.0 (the "License"); you may not use this software except in compliance with the License. You may obtain a copy of the License at

http://www.apache.org/licenses/LICENSE-2.0

Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License.

Contributor License Agreement

This project requires contributors to accept the terms in the Contributor License Agreement (CLA).

Please note that contributors to the NGT repository on GitHub (https://github.com/yahoojapan/NGT) shall be deemed to have accepted the CLA without individual written agreements.

Contact Person

masajiro

Publications

ONNG

• Iwasaki, M., Miyazaki, D.: Optimization of Indexing Based on k-Nearest Neighbor Graph for Proximity. arXiv:1810.07355 [cs] (2018). (pdf)

PANNG

• Iwasaki, M.: Pruned Bi-directed K-nearest Neighbor Graph for Proximity Search. Proc. of SISAP2016 (2016) 20-33. (pdf)
• Sugawara, K., Kobayashi, H. and Iwasaki, M.: On Approximately Searching for Similar Word Embeddings. Proc. of ACL2016 (2016) 2265-2275. (pdf)

ANNGT

• Iwasaki, M.: Applying a Graph-Structured Index to Product Image Search (in Japanese). IIEEJ Journal 42(5) (2013) 633-641. (pdf)
• Iwasaki, M.: Proximity search using approximate k nearest neighbor graph with a tree structured index (in Japanese). IPSJ Journal 52(2) (2011) 817-828. (pdf)

ANNG

• Iwasaki, M.: Proximity search in metric spaces using approximate k nearest neighbor graph (in Japanese). IPSJ Trans. on Database 3(1) (2010) 18-28. (pdf)