A locality sensitive hashing library with an emphasis on large, highly-dimensional datasets.
- Fast and memory-efficient calculations using sparse matrices.
- Built-in support for key-value storage backends: pure-Python, Redis (memory-bound), LevelDB, BerkeleyDB
- Multiple hash indexes support (based on Kay Zhu's lshash)
- Built-in support for common distance/objective functions for ranking outputs.
SparseLSH is based on a fork of Kay Zhu's lshash, and is suited for datasets that won't fit into main memory or are highly dimensional. Using sparse matrices allows for speedups of easily over an order of magnitude compared to using dense, list-based or numpy array-based vector math. Sparse matrices also makes it possible to deal with these datasets purely in memory using python dicts or through Redis. When this isn't appropriate, you can use one of the disk-based key-value stores, LevelDB and BerkeleyDB. Serialization is done using cPickle (for raw C speedups), falling back to python pickle if it's not available.
The easy way:
pip install sparselsh
Or you can clone this repo and follow these instructions:
SparseLSH
depends on the following libraries:
Optionally (for in-memory and disk-based persistence):
To install (minimal install):
python setup.py install
If you would like to use the LevelDB or Redis storage backends, you can
install the dependencies from the optional-requirements.txt
:
pip install -r optional-requirements.txt
You can quickly use LSH using the bundled sparselsh
command line utility. Simply pass the path to a file containing records to be clustered, one per line, and the script will output groups of similar items.
sparselsh path/to/recordsfile.txt
To create 4-bit hashes for input data of 7 dimensions:
from sparselsh import LSH
from scipy.sparse import csr_matrix
X = csr_matrix([
[3, 0, 0, 0, 0, 0, -1],
[0, 1, 0, 0, 0, 0, 1],
[1, 1, 1, 1, 1, 1, 1]
])
# One label for each input point
y = ["label-one", "second", "last"]
lsh = LSH(
4,
X.shape[1],
num_hashtables=1,
storage_config={"dict":None}
)
# If you're using >= v2.1.0, this is much faster
# lsh.index(X, extra_data=y)
# For all versions
for ix in range(X.shape[0]):
x = X.getrow(ix)
c = y[ix]
lsh.index( x, extra_data=c)
# Build a 1-D (single row) sparse matrix
X_sim = csr_matrix([[1, 1, 1, 1, 1, 1, 0]])
# find the point in X nearest to X_sim
points = lsh.query(X_sim, num_results=1)
The query above result in a list of matrix-class tuple & similarity
score tuples. A lower score is better in this case (the score here is 1.0).
Here's a breakdown of the return value when query
is called with a
single input row (1-dimensional sparse matrix, the output is different
when passing multiple query points):
[((<1x7 sparse matrix of type '<type 'numpy.int64'>' with 7 stored elements in Compressed Sparse Row format>, 'label'), 1.0)]
We can look at the most similar matched item by accessing the sparse array
and invoking it's todense
function:
# ,------- Get first result-score tuple
# | ,---- Get data. [1] is distance score
# | | ,-- Get the point. [1] is extra_data
# | | |
In [11]: print points[0][0][0].todense()
[[1 1 1 1 1 1 1]]
You can also pass multiple records to query
by simply increasing the
dimension of the input to query
. This will change the output data
to have one extra dimension, representing the input query. (NOTE: When
then dimension is 1, a.k.a. a single sparse row, this extra dimension won't
be added.) Here's the output when query
is passed a 2-dimensional input:
[
[((<1x7 sparse matrix ...>, 'label'), 1.0)],
[((<1x7 sparse matrix ...>, 'extra'), 0.8),
((<1x7 sparse matrix ...>, 'another'), 0.3)]
]
Here, you can see an extra dimension, one for each query point passed
to query
. The data structure for each query point result is the same
as the 1-Dimensional output.
Most of the parameters are supplied at class init time:
LSH( hash_size,
input_dim,
num_of_hashtables=1,
storage_config=None,
matrices_filename=None,
overwrite=False)
Parameters:
hash_size:
The length of the resulting binary hash. This controls how many "buckets"
there will be for items to be sorted into.
input_dim:
The dimension of the input vector. This needs to be the same as the input
points.
num_hashtables = 1:
(optional) The number of hash tables used. More hashtables increases the
probability of hash-collisions and the more similar items are likely
to be found for a query item. Increase this to get better accuracy
at the expense of increased memory usage.
storage = None:
(optional) A dict representing the storage backend and configuration
options. The following storage backends are supported with the following
configurations:
In-Memory Python Dictionary:
{"dict": None} # Takes no options
Redis:
{"redis": {"host": "127.0.0.1", "port": 6379, "db": 0}
LevelDB:
{"leveldb":{"db": "ldb"}}
Where "ldb" specifies the directory to store the LevelDB database.
(In this case it will be `./ldb/`)
Berkeley DB:
{"berkeleydb":{"filename": "./db"}}
Where "filename" is the location of the database file.
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. If you change the input
dimensions or the number of hashtables, you'll need to set the following
option, overwrite, to True, or delete this file.
overwrite = False:
(optional) Whether to overwrite the matrices file if it already exists.
-
To index a data point of a given
LSH
instance: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.
This can be used to hold values like class labels, URIs, titles, etc.
This function returns nothing.
To query a data point against a given LSH
instance:
lsh.query(query_point, num_results=None, distance_func="euclidean")
Parameters:
query_point:
The query data point is a sparse CSR matrix.
num_results = None:
(optional) Integer, specifies the max amount of results to be
returned. If not specified all candidates will be returned as a
list in ranked order.
NOTE: You do not save processing by limiting the results. Currently,
a similarity ranking and sort is done on all items in the hashtable
regardless if this parameter.
distance_func = "euclidean":
(optional) Distance function to use to rank the candidates. By default
euclidean distance function will be used.
Returns a list of tuples, each of which has the original input point (which will be a tuple of csr-matrix, extra_data or just the csr-matrix if no extra data was supplied) and a similarity score.