This repository contains a unified pipeline with seven fine-grained components for performing graph-based ANNS algorithms. Given a graph-based ANNS algorithms (such as NSG, HNSW), we can break it down to seven fine-grained components according to this framework, and run it through the CGraph's scheduling. Interestingly, we can obtain a better algorithm quickly by combining the existing algorithms' best-performing components, even without designing a new component. Of course, we can also optimize the components to further improve the algorithm's performance.
Approximate Nearest Neighbor Search(ANNS) is a task that finds the approximate nearest neighbors among a high-dimensional dataset for a query via a well-designed index. According to the index adopted, the existing ANNS algorithms can be divided into four major types: hashing-based; tree-based; quantization-based; and graph-based algorithms. Recently, graph-based algorithms have emerged as a highly effective option for ANNS. Thanks to graph-based ANNS algorithms' extraordinary ability to express neighbor relationships, they only need to evaluate fewer points of dataset to receive more accurate results, comparing to other categories.
This project supports MacOS, Linux, and Windows platforms, requiring CGraph. You can download this project by running the following instructions.
git clone --recursive [email protected]:whenever5225/GraphANNS.git
If you are a developer using CLion as the IDE, you can open the CMakeLists.txt file as a project for compiling.
You need to configure the data file, index file, and parameters in here (please refer to here for test data download). Suppose you use 'npg' algorithm and 'siftsmall' dataset, you can modify the data path to yours in its configure file (graph_anns_define.h):
const static char* GA_ALG_BASE_PATH = "your_base_data_root_path/siftsmall/siftsmall_base.fvecs";
const static char* GA_ALG_QUERY_PATH = "your_base_data_root_path/siftsmall/siftsmall_query.fvecs";
const static char* GA_ALG_GROUNDTRUTH_PATH = "your_base_data_root_path/siftsmall/siftsmall_groundtruth.ivecs";
const static char* GA_ALG_INDEX_PATH = "your_base_data_root_path/anns.index";
const static unsigned GA_NPG_L_CANDIDATE = 100; // size of candidate set for neighbor selection
const static unsigned GA_NPG_R_NEIGHBOR = 100; // size of neighbor set
const static unsigned GA_NPG_C_NEIGHBOR = 200; // number of visited candidate neighbors when neighbor selection
const static unsigned GA_NPG_K_INIT_GRAPH = 20; // number of neighbors of initial graphHINT: Please refer here for the description of fvecs/ivecs format.