Part of the code for the paper Practical Hash Functions for Similarity Estimation and Dimensionality Reduction at NIPS'17
Switch branches/tags
Nothing to show
Clone or download
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Permalink
Failed to load latest commit information.
src
.gitignore
LICENSE
README.md
poster.pdf

README.md

Nips_MT

This is the code for the paper "Practical Hash Functions for Similarity Estimation and Dimensionality Reduction" at NIPS'17

The code is written in C++ and contains reference implementations of several hash functions including Mixed Tabulation [1], Simple Tabulation [2], and Twisted Tabulation [3] as well as implementation of all experiments in the paper except the LSH experiments. The code also contains python scripts for generating plots from the data.

Overview

The code is split into several parts all contained in the src folder:

  • src/framework contains the implementations of different hash functions and sketches.
  • src/plotGenerator contains python scripts for generating plots once the experiments have been run. These require Numpy and Matplotlib to work.
  • src/ contains the code for the experiments as well as a Makefile to compile them. It also contains the code to generate the synthetic datasets.

Before you try to run the code plese read the sections below on randomness and data.

Also, for the code to work, you need to create a folder named "output" in the src/ and the src/plotGenerator/ folders.

Randomness

The code provided in this repo uses C++11 random generation as standard for portability. This is not intended!! We strongly recommend that any user downloads a large seed of random bytes from e.g. random.org and uses this instead. There is information in src/framework/hashing.h, src/framework/hashing_more.h, and src/framework/sketches.h on how to use such a random seed instead.

Data

The news20 and MNIST data sets are not included in this repo. They can be downloaded here: News20 and MNIST

To use the code, place the data files in the src/data/ folder and compile and run the code of the news20_change_format.cpp file to create a data file with the news20 dataset represented as indicator vectors.

External software

The code uses the official implementations of MurmurHash3, CityHash, and BLAKE2 and wraps these in a class similar to the reference implementations of the other hash functions.

The official implementations of these hash functions are available at (and also included in the src/framework folder):

References

[1] Søren Dahlgaard, Mathias Bæk Tejs Knudsen, Eva Rotenberg, Mikkel Thorup: Hashing for Statistics over K-Partitions. FOCS 2015: 1292-1310

[2] Mihai Patrascu, Mikkel Thorup: The Power of Simple Tabulation Hashing. J. ACM 59(3): 14:1-14:50 (2012)

[3] Mihai Patrascu, Mikkel Thorup: Twisted Tabulation Hashing. SODA 2013: 209-228