cgtuebingen / product-quantization-tree

GPU-based large scale Approx. Nearest Neighbor Search, accepted at CVPR 2016

Efficient Large-scale Approximate Nearest Neighbor Search on the GPU

This repository contains the implementation of Product Quantization Trees (PQT) for large scale nearest neighbor search.

@inproceedings{PQT,
    author = "Patrick Wieschollek and Oliver Wang and Alexander Sorkine-Hornung and Hendrik P.A. Lensch",
    title = "Efficient Large-scale Approximate Nearest Neighbor Search on the GPU",
    booktitle = "IEEE Conference on Computer Vision and Pattern Recognition (CVPR)",
    pages = "",
    month = "June",
    year = "2016",
    url = "goo.gl/4Zl5xB"
}

This experimental implementation is the first running on the GPU outperforming previous CPU approaches.

Usage

Download

Just by

git clone --recursive https://github.com/cgtuebingen/Product-Quantization-Tree.git

Requirements

To build this project you need

• cmake
• A C++11 capable compiler
• The CUDA toolkit and the CUDA samples.
• GFlags and GoogleTesting (as git submodules)

We use Toolkit v7.5, gcc 4.8.4 on Ubuntu 14.04 for Nvidia Titan X.

Preprocessing

To handle datasets and efficiently read them from (ram-)disk we converted the official datasets from SIFT1M, SIFT1B in our own format:

Each *.umem, *.imem, *.fmem has the following layout (header and content)

uint number_of_vectors
uint dimen_of_vector
... header are 20 bytes, next data start at byte 20:
T consecutive array of data, each entry is a T

Examples

* .umem: `uint  uint 0 0 ... 0 0 uint8_t uint8_t uint8_t uint8_t uint8_t uint8_t ...  uint8_t`
* .imem: `uint  uint 0 0 ... 0 0 int int int int int int ...  int`
* .fmem: `uint  uint 0 0 ... 0 0 float float float float float float ...  float`

We provide script to convert these datasets

./convert_fvecs --fvecs src/path/to/db.fvecs --umem dst/path/to/db.umem
./convert_bvecs --bvecs src/path/to/db.bvecs --umem dst/path/to/db.umem
./convert_ivecs --ivecs src/path/to/db.ivecs --imem dst/path/to/db.imem

Query (offline phase)

Building the index structure (Product-Quantization-Tree) is done by tool_createdb. This creates the index structure, prepares the database and dump all intermediate values into binary files.

Possible Flags (see ./tool_createdb -h):

- device    "selected cuda device"
- c1        "number of clusters in first level"
- c2        "number of refinements in second level"
- p         "parts per vector"
- dim       "expected dimension for each vector"
- lineparts "vectorparts for reranking informations"
- chunksize "number of vectors per chunk"
- hashsize  "maximal number of bins"
- basename  "prefix for generated data"
- dataset   "patch to vector dataset"

Query (online phase)

The query process is done batchwise using tool_query. The accompanying example return the best and second best found vector. For possible Flags see ./tool_query -h.