Header-only C++ HNSW implementation with python bindings, insertions and updates.
applied to KNN graph construction
NEWS:
version 0.7.0
- Added support to filtering (#402, #430) by @kishorenc
- Added python interface for filtering (though note its performance is limited by GIL) (#417) by @gtsoukas
- Added support for replacing the elements that were marked as delete with newly inserted elements (to control the size of the index, #418) by @dyashuni
- Fixed data races/deadlocks in updates/insertion, added stress test for multithreaded operation (#418) by @dyashuni
- Documentation, tests, exception handling, refactoring (#375, #379, #380, #395, #396, #401, #406, #404, #409, #410, #416, #415, #431, #432, #433) by @jlmelville, @dyashuni, @kishorenc, @korzhenevski, @yoshoku, @jianshu93, @PLNech
- global linkages (#383) by @MasterAler, USE_SSE usage in MSVC (#408) by @alxvth
- Lightweight, header-only, no dependencies other than C++ 11
- Interfaces for C++, Python, external support for Java and R (https://github.com/jlmelville/rcpphnsw).
- Has full support for incremental index construction and updating the elements. Has support for element deletions (by marking them in index). Index is picklable.
- Can work with custom user defined distances (C++).
- Significantly less memory footprint and faster build time compared to current nmslib's implementation.
Description of the algorithm parameters can be found in ALGO_PARAMS.md.
Distance | parameter | Equation |
---|---|---|
Squared L2 | 'l2' | d = sum((Ai-Bi)^2) |
Inner product | 'ip' | d = 1.0 - sum(Ai*Bi) |
Cosine similarity | 'cosine' | d = 1.0 - sum(Ai*Bi) / sqrt(sum(Ai*Ai) * sum(Bi*Bi)) |
Note that inner product is not an actual metric. An element can be closer to some other element than to itself. That allows some speedup if you remove all elements that are not the closest to themselves from the index.
For other spaces use the nmslib library https://github.com/nmslib/nmslib.
hnswlib.Index(space, dim)
creates a non-initialized index an HNSW in spacespace
with integer dimensiondim
.
hnswlib.Index
methods:
-
init_index(max_elements, M = 16, ef_construction = 200, random_seed = 100, allow_replace_deleted = False)
initializes the index from with no elements.max_elements
defines the maximum number of elements that can be stored in the structure(can be increased/shrunk).ef_construction
defines a construction time/accuracy trade-off (see ALGO_PARAMS.md).M
defines tha maximum number of outgoing connections in the graph (ALGO_PARAMS.md).allow_replace_deleted
enables replacing of deleted elements with new added ones.
-
add_items(data, ids, num_threads = -1, replace_deleted = False)
- inserts thedata
(numpy array of vectors, shape:N*dim
) into the structure.num_threads
sets the number of cpu threads to use (-1 means use default).ids
are optional N-size numpy array of integer labels for all elements indata
.- If index already has the elements with the same labels, their features will be updated. Note that update procedure is slower than insertion of a new element, but more memory- and query-efficient.
replace_deleted
replaces deleted elements. Note it allows to save memory.- to use it
init_index
should be called withallow_replace_deleted=True
- to use it
- Thread-safe with other
add_items
calls, but not withknn_query
.
-
mark_deleted(label)
- marks the element as deleted, so it will be omitted from search results. Throws an exception if it is already deleted. -
unmark_deleted(label)
- unmarks the element as deleted, so it will be not be omitted from search results. -
resize_index(new_size)
- changes the maximum capacity of the index. Not thread safe withadd_items
andknn_query
. -
set_ef(ef)
- sets the query time accuracy/speed trade-off, defined by theef
parameter ( ALGO_PARAMS.md). Note that the parameter is currently not saved along with the index, so you need to set it manually after loading. -
knn_query(data, k = 1, num_threads = -1, filter = None)
make a batch query fork
closest elements for each element of thedata
(shape:N*dim
). Returns a numpy array of (shape:N*k
).num_threads
sets the number of cpu threads to use (-1 means use default).filter
filters elements by its labels, returns elements with allowed ids. Note that search with a filter works slow in python in multithreaded mode. It is recommended to setnum_threads=1
- Thread-safe with other
knn_query
calls, but not withadd_items
.
-
load_index(path_to_index, max_elements = 0, allow_replace_deleted = False)
loads the index from persistence to the uninitialized index.max_elements
(optional) resets the maximum number of elements in the structure.allow_replace_deleted
specifies whether the index being loaded has enabled replacing of deleted elements.
-
save_index(path_to_index)
saves the index from persistence. -
set_num_threads(num_threads)
set the default number of cpu threads used during data insertion/querying. -
get_items(ids)
- returns a numpy array (shape:N*dim
) of vectors that have integer identifiers specified inids
numpy vector (shape:N
). Note that for cosine similarity it currently returns normalized vectors. -
get_ids_list()
- returns a list of all elements' ids. -
get_max_elements()
- returns the current capacity of the index -
get_current_count()
- returns the current number of element stored in the index
Read-only properties of hnswlib.Index
class:
-
space
- name of the space (can be one of "l2", "ip", or "cosine"). -
dim
- dimensionality of the space. -
M
- parameter that defines the maximum number of outgoing connections in the graph. -
ef_construction
- parameter that controls speed/accuracy trade-off during the index construction. -
max_elements
- current capacity of the index. Equivalent top.get_max_elements()
. -
element_count
- number of items in the index. Equivalent top.get_current_count()
.
Properties of hnswlib.Index
that support reading and writing:
-
ef
- parameter controlling query time/accuracy trade-off. -
num_threads
- default number of threads to use inadd_items
orknn_query
. Note that callingp.set_num_threads(3)
is equivalent top.num_threads=3
.
- Creating index, inserting elements, searching, serialization/deserialization
- Filtering during the search with a boolean function
- Deleting the elements and reusing the memory of the deleted elements for newly added elements
An example of creating index, inserting elements, searching and pickle serialization:
import hnswlib
import numpy as np
import pickle
dim = 128
num_elements = 10000
# Generating sample data
data = np.float32(np.random.random((num_elements, dim)))
ids = np.arange(num_elements)
# Declaring index
p = hnswlib.Index(space = 'l2', dim = dim) # possible options are l2, cosine or ip
# Initializing index - the maximum number of elements should be known beforehand
p.init_index(max_elements = num_elements, ef_construction = 200, M = 16)
# Element insertion (can be called several times):
p.add_items(data, ids)
# Controlling the recall by setting ef:
p.set_ef(50) # ef should always be > k
# Query dataset, k - number of the closest elements (returns 2 numpy arrays)
labels, distances = p.knn_query(data, k = 1)
# Index objects support pickling
# WARNING: serialization via pickle.dumps(p) or p.__getstate__() is NOT thread-safe with p.add_items method!
# Note: ef parameter is included in serialization; random number generator is initialized with random_seed on Index load
p_copy = pickle.loads(pickle.dumps(p)) # creates a copy of index p using pickle round-trip
### Index parameters are exposed as class properties:
print(f"Parameters passed to constructor: space={p_copy.space}, dim={p_copy.dim}")
print(f"Index construction: M={p_copy.M}, ef_construction={p_copy.ef_construction}")
print(f"Index size is {p_copy.element_count} and index capacity is {p_copy.max_elements}")
print(f"Search speed/quality trade-off parameter: ef={p_copy.ef}")
An example with updates after serialization/deserialization:
import hnswlib
import numpy as np
dim = 16
num_elements = 10000
# Generating sample data
data = np.float32(np.random.random((num_elements, dim)))
# We split the data in two batches:
data1 = data[:num_elements // 2]
data2 = data[num_elements // 2:]
# Declaring index
p = hnswlib.Index(space='l2', dim=dim) # possible options are l2, cosine or ip
# Initializing index
# max_elements - the maximum number of elements (capacity). Will throw an exception if exceeded
# during insertion of an element.
# The capacity can be increased by saving/loading the index, see below.
#
# ef_construction - controls index search speed/build speed tradeoff
#
# M - is tightly connected with internal dimensionality of the data. Strongly affects memory consumption (~M)
# Higher M leads to higher accuracy/run_time at fixed ef/efConstruction
p.init_index(max_elements=num_elements//2, ef_construction=100, M=16)
# Controlling the recall by setting ef:
# higher ef leads to better accuracy, but slower search
p.set_ef(10)
# Set number of threads used during batch search/construction
# By default using all available cores
p.set_num_threads(4)
print("Adding first batch of %d elements" % (len(data1)))
p.add_items(data1)
# Query the elements for themselves and measure recall:
labels, distances = p.knn_query(data1, k=1)
print("Recall for the first batch:", np.mean(labels.reshape(-1) == np.arange(len(data1))), "\n")
# Serializing and deleting the index:
index_path='first_half.bin'
print("Saving index to '%s'" % index_path)
p.save_index("first_half.bin")
del p
# Re-initializing, loading the index
p = hnswlib.Index(space='l2', dim=dim) # the space can be changed - keeps the data, alters the distance function.
print("\nLoading index from 'first_half.bin'\n")
# Increase the total capacity (max_elements), so that it will handle the new data
p.load_index("first_half.bin", max_elements = num_elements)
print("Adding the second batch of %d elements" % (len(data2)))
p.add_items(data2)
# Query the elements for themselves and measure recall:
labels, distances = p.knn_query(data, k=1)
print("Recall for two batches:", np.mean(labels.reshape(-1) == np.arange(len(data))), "\n")
- creating index, inserting elements, searching, serialization/deserialization
- filtering during the search with a boolean function
- deleting the elements and reusing the memory of the deleted elements for newly added elements
- multithreaded usage
You can install from sources:
apt-get install -y python-setuptools python-pip
git clone https://github.com/nmslib/hnswlib.git
cd hnswlib
pip install .
or you can install via pip:
pip install hnswlib
Contributions are highly welcome!
Please make pull requests against the develop
branch.
When making changes please run tests (and please add a test to tests/python
in case there is new functionality):
python -m unittest discover --start-directory tests/python --pattern "bindings_test*.py"
- Non-metric space library (nmslib) - main library(python, C++), supports exotic distances: https://github.com/nmslib/nmslib
- Faiss library by facebook, uses own HNSW implementation for coarse quantization (python, C++): https://github.com/facebookresearch/faiss
- Code for the paper "Revisiting the Inverted Indices for Billion-Scale Approximate Nearest Neighbors" (current state-of-the-art in compressed indexes, C++): https://github.com/dbaranchuk/ivf-hnsw
- Amazon PECOS https://github.com/amzn/pecos
- TOROS N2 (python, C++): https://github.com/kakao/n2
- Online HNSW (C++): https://github.com/andrusha97/online-hnsw)
- Go implementation: https://github.com/Bithack/go-hnsw
- Python implementation (as a part of the clustering code by by Matteo Dell'Amico): https://github.com/matteodellamico/flexible-clustering
- Julia implmentation https://github.com/JuliaNeighbors/HNSW.jl
- Java implementation: https://github.com/jelmerk/hnswlib
- Java bindings using Java Native Access: https://github.com/stepstone-tech/hnswlib-jna
- .Net implementation: https://github.com/curiosity-ai/hnsw-sharp
- CUDA implementation: https://github.com/js1010/cuhnsw
- Rust implementation https://github.com/rust-cv/hnsw
- Rust implementation for memory and thread safety purposes and There is A Trait to enable the user to implement its own distances. It takes as data slices of types T satisfying T:Serialize+Clone+Send+Sync.: https://github.com/jean-pierreBoth/hnswlib-rs
To download and extract the bigann dataset (from root directory):
python tests/cpp/download_bigann.py
To compile:
mkdir build
cd build
cmake ..
make all
To run the test on 200M SIFT subset:
./main
The size of the BigANN subset (in millions) is controlled by the variable subset_size_millions hardcoded in sift_1b.cpp.
To generate testing data (from root directory):
cd tests/cpp
python update_gen_data.py
To compile (from root directory):
mkdir build
cd build
cmake ..
make
To run test without updates (from build
directory)
./test_updates
To run test with updates (from build
directory)
./test_updates update
- Visual search engine for 1M amazon products (MXNet + HNSW): website, code, demo by @ThomasDelteil
HNSW paper:
@article{malkov2018efficient,
title={Efficient and robust approximate nearest neighbor search using hierarchical navigable small world graphs},
author={Malkov, Yu A and Yashunin, Dmitry A},
journal={IEEE transactions on pattern analysis and machine intelligence},
volume={42},
number={4},
pages={824--836},
year={2018},
publisher={IEEE}
}
The update algorithm supported in this repository is to be published in "Dynamic Updates For HNSW, Hierarchical Navigable Small World Graphs" US Patent 15/929,802 by Apoorv Sharma, Abhishek Tayal and Yury Malkov.