Implementation of the Hamming Weight Tree in C++.
HWT supports two fundumental queries in binary datasets: 1) K nearest neighbor search 2) Insertion of new binary codes
Compiling and building the project requires cmake, make, hdf5 and hdf5-dev. To compile go to hwt folder and run cmake:
cmake CMakeLists.txt
make
This should create two executable files 1) hwt, for running the Hamming weight tree, and 2) linscan for running linear scan
hwt provides an implementaiton of the Hamming weight tree which supports fast Hamming nearest neighbor search as well as fast insertion of new items to the data strucutre
./hwt <.mat input file> <h5 output file> -K <#neighbors> -B <length of code> -Q <#queries> -T <capacity>
e.g.:/hwt binary.mat out.h5 -K 10 -B 64 -Q 100 -T 1000
The input file should be a .mat file containing the matrix of binary codes B and the matrix of query points Q. In both matrices the columns reporsent the data points and rows represent the bits. To reduce the storage cost, each 8 bits of binary code must be compact into a byte, ranging from 0 to 255. Therefore, the number of rows is equal to (length of code)/8. K is the number of nearest neighbors to retreive, B is the legnth of codes, Q is the number of queries and T is the maximum number of nodes that can assigned to a leaf node.
linscan provides an implementation of the linear search to find the K nearest neighbors for dynamic datasets.
./linscan <.mat input file> <h5 output file> -K <#neighbors> -B <length of code> -Q <#queries>
e.g.:/hwt binary.mat out.h5 -K 10 -B 64 -Q 100
The syntax is similar to hwt except that it does not require to specify the capacity.