ruby-lsh

An implementation of Locality-Sensitive Hashing in Ruby.

Locality-Sensitive Hashing is an efficient technique for finding nearest neighbors in high-dimensional spaces.

Uses JBlas on JRuby and GSL on any other Ruby. Supports two backends (in-memory and Redis-backed) and includes a small HTTP API.

Usage

See examples/evaluation.rb

require 'lsh'
index = LSH::Index.new({
  :dim => 100, 
  :number_of_random_vectors => 8, 
  :window => Float::INFINITY, 
  :number_of_independent_projections => 150
}) # Creates an in-memory binary LSH index for 100-dim vectors, 8 bits, 150 independent projections
v1 = index.random_vector(100)
v2 = index.random_vector(100)
v3 = index.random_vector(100) # Creating three random vectors
index.add v1
index.add v2
index.add v3 # Adding the three vectors to the index
index.query(v1) # Query the index for vectors that fall in the same LSH bucket as v1
index.query(v2, 1) # Query the index for vectors that fall in the same LSH bucket as v2, and in buckets at hamming distance 1 of that bucket

Using the Redis backend

By default, the LSH index will be stored in memory. A Redis-backed storage is also available, and can be constructed as follows:

storage = LSH::Storage::RedisBackend.new
index = LSH::Index.new({
  :dim => 100,
  :number_of_random_vectors => 8,
  :window => Float::Infinity,
  :number_of_independent_projections => 150
}, storage)

Once created, the index can then be reused:

storage = LSH::Storage::RedisBackend.new
index = LSH::Index.new(storage.parameters, storage) if storage.has_index?

This will connect to a Redis backend on localhost and store binary dumps of the vectors (including the projections) in a 'data' directory, caching vectors in RAM to speed up comparison. This can be overridden as follows:

storage = LSH::Storage::RedisBackend.new(:redis => { :host => '127.0.0.1', :port => 6379 },
                                         :data_dir => 'data',
                                         :cache_vectors => TRUE)

The Redis-backed LSH index is faster using the MRI than JRuby, due to the time it takes to load vectors from their binary representations on disk. GSL is much faster than JBLAS on that point.

Using the Web frontend

This gem includes a minimal Web API, built using Sinatra. See examples/config.ru for an example setup.

$ cd examples
$ rackup
$ curl --data-urlencode data@vector.json http://localhost:9292/index # Adds a vector to the index
$ curl --data-urlencode data@vector.json http://localhost:9292/query # Query the index

Or you can associate your vectors with ids and query using them.

$ cd examples
$ rackup
$ curl --data-urlencode data@vector.json -d'id=foo' http://localhost:9292/index # Adds a vector with id 'foo' to the index
$ curl -d'id=foo' http://localhost:9292/query-ids # Query the index

GSL notes

If you get a compilation error when installing GSL, try this version:

https://github.com/romanbsd/rb-gsl

As you will need a version of GSL that includes this patch:

https://gist.github.com/1217974

Licensing terms and authorship

See 'COPYING' and 'AUTHORS' files.

lsh