PANN: Parallel Approximate Nearest Neighbour Search Over Graphs

Prerequisites

We have tested PANN on the following versions of prerequisites.

g++ 5.4.0, 9.3.0
CUDA 10.0, 10.1

Our referenced paper homepage: SONG.

Dataset

• https://www.csie.ntu.edu.tw/~cjlin/libsvmtools/datasets/multiclass.html
• http://corpus-texmex.irisa.fr/

Several Settings

We have two settings as s1 and s2. You can change the code in ./main.cu to run under different settings.

Run

You can run ./run.sh to generate the result.

• ./generate_template.sh is to generate a copy of all the files.
• ./fill_parameters.sh q dimension l2
  • For the parameter q, it refers to the length of topk when building the graph.
  • For the parameter dimension, it means the dimension of the dataset.
• ./build_graph.sh cifar10.txt num dimension l2
  • For the parameter num, it refers to the number of the points of the dataset.
  • For the parameter dimension, it means the dimension of the dataset.
• ./test_query.sh cifar10.txt num l2 s
  • For the parameter num, it refers to the number of the points of the dataset.
  • For the parameter s, it refers to the length of the candidate queue when querying.

Evaluate

You can run ./evaluate.sh to evalute the recall.