LSHash

A fast Python implementation of locality sensitive hashing with persistance support.

Based on original source code https://github.com/kayzhu/LSHash

Highlights

• Python3 support
• Load & save hash tables to local disk
• Fast hash calculation for large amount of high dimensional data through the use of numpy arrays.
• Built-in support for persistency through Redis.
• Multiple hash indexes support.
• Built-in support for common distance/objective functions for ranking outputs.

Installation

LSHash depends on the following libraries:

• numpy
• bitarray (if hamming distance is used as distance function)

Optional - redis (if persistency through Redis is needed)

To install from sources:

$ git clone https://github.com/loretoparisi/lshash.git
$ python setup.py install

To install from PyPI:

$ pip install lshashpy3
$ python -c "import lshashpy3 as lshash; print(lshash.__version__);"

Quickstart

To create 6-bit hashes for input data of 8 dimensions:

# create 6-bit hashes for input data of 8 dimensions:
lsh = LSHash(6, 8)

# index vector
lsh.index([2,3,4,5,6,7,8,9])

# index vector and extra data
lsh.index([10,12,99,1,5,31,2,3], extra_data="vec1")
lsh.index([10,11,94,1,4,31,2,3], extra_data="vec2")

# query a data point
top_n = 1
nn = lsh.query([1,2,3,4,5,6,7,7], num_results=top_n, distance_func="euclidean")
print(nn)

# unpack vector, extra data and vectorial distance
top_n = 3
nn = lsh.query([10,12,99,1,5,30,1,1], num_results=top_n, distance_func="euclidean")
   for ((vec,extra_data),distance) in nn:
       print(vec, extra_data, distance)

To save hash table to disk:

lsh = LSHash(hash_size=k, input_dim=d, num_hashtables=L,
    storage_config={ 'dict': None },
    matrices_filename='weights.npz',
    hashtable_filename='hash.npz',
    overwrite=True)

lsh.index([10,12,99,1,5,31,2,3], extra_data="vec1")
lsh.index([10,11,94,1,4,31,2,3], extra_data="vec2")
lsh.save()

To load hash table from disk and perform a query:

lsh = LSHash(hash_size=k, input_dim=d, num_hashtables=L,
    storage_config={ 'dict': None },
    matrices_filename='weights.npz',
    hashtable_filename='hash.npz',
    overwrite=True)

top_n = 3
nn = lsh.query([10,12,99,1,5,30,1,1], num_results=top_n, distance_func="euclidean")
print(nn)

API

• To initialize a LSHash instance:

k = 6 # hash size
L = 5  # number of tables
d = 8 # Dimension of Feature vector
LSHash(hash_size=k, input_dim=d, num_hashtables=L,
   storage_config={ 'dict': None },
   matrices_filename='weights.npz',
   hashtable_filename='hash.npz',
   overwrite=True)

parameters:

hash_size:

The length of the resulting binary hash.

input_dim:

The dimension of the input vector.

num_hashtables = 1:

(optional) The number of hash tables used for multiple lookups.

storage = None:

(optional) Specify the name of the storage to be used for the index storage. Options include "redis".

matrices_filename = None:

(optional) Specify the path to the .npz file random matrices are stored or to be stored if the file does not exist yet

hashtable_filename = None:

(optional) Specify the path to the .npz file hash table are stored or to be stored if the file does not exist yet

overwrite = False:

(optional) Whether to overwrite the matrices file if it already exist

• To index a data point of a given LSHash instance, e.g., lsh:

lsh.index(input_point, extra_data=None):

parameters:

input_point:

The input data point is an array or tuple of numbers of input_dim.

extra_data = None:

(optional) Extra data to be added along with the input_point.

• To query a data point against a given LSHash instance, e.g., lsh:

lsh.query(query_point, num_results=None, distance_func="euclidean"):

parameters:

query_point:

The query data point is an array or tuple of numbers of input_dim.

num_results = None:

(optional) The number of query results to return in ranked order. By default all results will be returned.

distance_func = "euclidean":

(optional) Distance function to use to rank the candidates. By default "euclidean" distance function will be used. Distance function can be "euclidean", "true_euclidean", "centred_euclidean", "cosine", "l1norm".

• To save the hash table currently indexed:

lsh.save():