GQR

A General and Efficient Querying Method for Learning to Hash (SIGMOD 2018)

---

Why GQR framework for similarity search

• Fast Training (Matlab and Automatic Parallelization)
• Support of Multiple Hash Tables (Like k-nn Graph supports k > 1)

Dependences

• CMake
• Matlab

Run

-Step 1: clone the repository

git clone --recursive https://github.com/lijinf2/gqr.git

cd gqr

-Step 2: train models, which are stored under folder hashingCodeTXT

cd learn/PCAH

mkdir hashingCodeTXT && bash run.sh

cd ../../

-Step 3: make app and run

mkdir build && cd ./build && cmake ../ -DCMAKE_BUILD_TYPE=Release

make search

cd ../script && bash search.sh

You may refer to folder ./script for detailed explanations and more instructions.

Reproduce results in the SIGMOD paper

cd reproduce

bash prepareData

./runExperiment.sh PCAH GQR 12 1 cifar60k

bash ./runMainExperiments

Support

Hashing Algorithms

• ITQ: Iterative quantization
• PCAH: PCA Hashing
• PCARR: PCA hashing with random rotation
• SH: Spectral Hashing
• SpH: Spherical Hashing
• KMH: KMeans Hashing
• IsoH: Isotropic Hashing
• SIM: Random Projecting Hashing.
• LMIP: Length Marked (variant length) Inner Product
  • only work with query method LM
  • based on SIM which generate random projecting bits, and extra bits is generated for representing NORM(Length)

Query Algorithms

• HR: Hamming Ranking
• HL: Hash Lookup
• GQR: Generate-to-probe quantization ranking
• LM: Length Marked ranking(work the LMIP)
  • use both random projecting bits and extra bis generated in LMIP to rank

Acknowledgement

GQR project is developed based on LSHBOX (https://github.com/RSIA-LIESMARS-WHU/LSHBOX) and MatlabFunc (https://github.com/dengcai78/MatlabFunc). Great appreciation to the contributors of LSHBOX and MatlabFunc.

More

Distribued GQR on general system for similarity search

Reference

A General and Efficient Querying Method for Learning to Hash (SIGMOD 2018)