Neighborhood Graph and Tree for Indexing High-dimensional Data
NGT provides commands and a library for performing high-speed approximate nearest neighbor searches against a large volume of data (several million to several 10 million items of data) in high dimensional vector data space (several ten to several thousand dimensions).
- 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)
$ 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 gcc@9 $ ln -s ./gcc-9 /usr/local/bin/gcc $ ln -s ./g++-9 /usr/local/bin/g++ $ export CXX=g++ $ export CC=gcc $ unzip NGT-x.x.x.zip $ cd NGT-x.x.x $ mkdir build $ cd build $ cmake .. $ make $ make install
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, please add the following parameter to improve the search time.
$ cmake -DNGT_LARGE_DATASET=ON ..
Supported Programming Languages
The followings are the results of ann benchmarks on an AWS c5.4xlarge instance for NGT v1.7.5.
Copyright (C) 2015-2019 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
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.
Yahoo Japan Corporation has aquired several patents on the technologies utilized in this software. However, the patent rights shall not be exercised under Apache License Version 2.0, only when the patented techniques are used with this software.
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.
- Iwasaki, M., Miyazaki, D.: Optimization of Indexing Based on k-Nearest Neighbor Graph for Proximity. arXiv:1810.07355 [cs] (2018). (pdf)
- 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)
- 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)
- 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)
Copyright © 2015-2019 Yahoo Japan Corporation All Rights Reserved.