This project works with Python on version 2.7+ and 3.5+, and on Linux, OSX and the Windows operating systems. To install:
pip install nmslib
To install from sources:
pip install --no-binary :all: nmslib
To install from sources, you may need to install Python dev-files. On Ubuntu, you can do it as follows:
sudo apt-get install python3-dev
Building on Windows requires Visual Studio 2015, see this project for more information.
Here is a simple example, but we also have Python notebooks with more elaborate end-to-end examples, which include even computation of gold-standard data (for both sparse and dense space):
import nmslib
import numpy
# create a random matrix to index
data = numpy.random.randn(10000, 100).astype(numpy.float32)
# initialize a new index, using a HNSW index on Cosine Similarity
index = nmslib.init(method='hnsw', space='cosinesimil')
index.addDataPointBatch(data)
index.createIndex({'post': 2}, print_progress=True)
# query for the nearest neighbours of the first datapoint
ids, distances = index.knnQuery(data[0], k=10)
# get all nearest neighbours for all the datapoint
# using a pool of 4 threads to compute
neighbours = index.knnQueryBatch(data, k=10, num_threads=4)It is possible to save both indexes (for some of the search methods) and serialized data in the binary format for subsequent faster loading. Then, one does not need to call index.addDataPointBatch and index.createIndex. Instead, one can first call index.saveIndex(indexLocation, save_data=True) and then
index = nmslib.init(method='hnsw', space='cosinesimil')
index.loadIndex(indexLocal, load_data=True)
One catch though is that for spaces l2 and cosinesimil, HNSW's method saveIndex always saves its own copy of data. In this case, we say that HNSW saves an optimized version of the index. Thus, to avoid data duplication one can set parameters of save_data and load_data to false. Examples of doing so can be found in sample Python notebooks. Note, though, that the function getDistance will not work properly unless the data is reloaded (this is certainly a deficiency, but it is not easy to fix).
The basic parameter tuning/selection guidelines are available here.
NMSLIB produces quite a few informational messages. By default, they are not shown in Python. To enable debugging, one should use the following commands before importing the library:
import logging
logging.basicConfig(level=logging.DEBUG)
To enable extra methods like those provided by FALCONN and LSHKIT you need to follow an extra couple steps.
These methods require a development version of the following libraries: Boost, GNU scientific library, and Eigen3. To install on Ubuntu:
sudo apt-get install libboost-all-dev libgsl0-dev libeigen3-dev
Next clone the repository and build with the C++ files using CMake:
cd similarity_search
cmake . -DWITH_EXTRAS=1
make
cd ..
Finally build and install the python extension:
cd python_bindings
pip install -r requirements.txt
python setup.py install