RcppHNSW

Rcpp bindings for hnswlib.

Status

February 4 2024 RcppHNSW 0.6.0 is released to CRAN, supporting hnswlib version 0.8.0.

September 19 2023 RcppHNSW 0.5.0 is released to CRAN, supporting hnswlib version 0.7.0, a getItems method for returning the items used to build the index and some performance improvements if your data is already column-stored. Also, a small roxygen problem with the package documentation was fixed.

July 18 2022 RcppHNSW 0.4.1 is released. Unfortunately, there are valgrind problems with the version of hnswlib used in RcppHNSW 0.4.0, so that has been rolled back.

July 16 2022 RcppHNSW 0.4.0 is released. This release matches hnswlib version 0.6.2, but otherwise adds no new features. Some minor CRAN check NOTEs are fixed and there is also a minor license change: previously the license was GPLv3. From this version, it now supports GPLv3 or later.

September 6 2020 RcppHNSW 0.3.0 is now available on CRAN, with multi-threading support.

August 30 2020. Although not yet on CRAN, support for building and searching an index in parallel (via the n_threads function argument and setNumThreads object method) has been added to the current development version (available via devtools::install_github). Thanks to Dmitriy Selivanov for a lot of the work on this.

September 20 2019. RcppHNSW 0.2.0 is now available on CRAN, up to date with hnswlib at https://github.com/nmslib/hnswlib/commit/c5c38f0, with new methods: size, resizeIndex and markDeleted. Also, a bug that prevented searching with datasets smaller than k has been fixed. Thanks to Yuxing Liao for spotting that.

January 21 2019. RcppHNSW is now available on CRAN.

October 20 2018. By inserting some preprocessor symbols into hnswlib, these bindings no longer require a non-portable compiler flag and hence will pass R CMD CHECK without any warnings: previously you would be warned about -march=native. The price paid is not using specialized functions for the distance calculations that are architecture-specific. I have not checked how bad the performance hit is. The old settings remain in src/Makevars and src/Makevars.win (commented out), if you want to build the project from source directly. Otherwise, Release 0.0.0.9000 is the last version with the old behavior, which can be installed with something like:

devtools::install_github("jlmelville/[email protected]")

hnswlib

hnswlib is a header-only C++ library for finding approximate nearest neighbors (ANN) via Hierarchical Navigable Small Worlds (Yashunin and Malkov, 2016). It is part of the nmslib project.

The RcppHNSW Package

An R package that interfaces with hnswlib, taking enormous amounts of inspiration from Dirk Eddelbuettel's RcppAnnoy package which did the same for the Annoy ANN C++ library.

One difference is that I use roxygen2 to generate the man pages. The NAMESPACE is still built manually, however (I don't believe you can export the classes currently).

Installing

From CRAN:

install.packages("RcppHNSW")

Development versions from github:

devtools::install_github("jlmelville/RcppHNSW")

Function example

irism <- as.matrix(iris[, -5])

# function interface returns results for all rows in nr x k matrices
all_knn <- RcppHNSW::hnsw_knn(irism, k = 4, distance = "l2")
# other distance options: "euclidean", "cosine" and "ip" (inner product distance)

# for high-dimensional data you may see a speed-up if you store the data
# where each *column* is an item to be indexed and searched. Set byrow = TRUE
# for this.
# Admittedly, the iris dataset is *not* high-dimensional
iris_by_col <- t(irism)
all_knn <- RcppHNSW::hnsw_knn(iris_by_col, k = 4, distance = "l2", byrow = FALSE)

# process can be split into two steps, so you can build with one set of data
# and search with another
ann <- hnsw_build(irism[1:100, ])
iris_nn <- hnsw_search(irism[101:150, ], ann, k = 5)

Class Example

As noted in the "Do not use named parameters" section below, you should avoid using named parameters when using class methods. But I do use them in a few places below to document the name of the parameters the positional arguments refer to.

library(RcppHNSW)
data <- as.matrix(iris[, -5])

# Create a new index using the L2 (squared Euclidean) distance
# nr and nc are the number of rows and columns of the data to be added, respectively
# ef and M determines speed vs accuracy trade off
# You must specify the maximum number of items to add to the index when it
# is created. But you can increase this number: see the next example
M <- 16
ef <- 200
dim <- ncol(data)
nitems <- nrow(data)
ann <- new(HnswL2, dim, nitems, M, ef)

# Add items to index
for (i in 1:nitems) {
  ann$addItem(data[i, ])
}

# Find 4 nearest neighbors of row 1
# indexes are in res$item, distances in res$distance
# set include_distances = TRUE to get distances as well as index
res <- ann$getNNsList(data[1, ], k = 4, include_distances = TRUE)

# It's more efficient to use the batch methods if you have all the data you
# need at once
ann2 <- new(HnswL2, dim, nitems, M, ef)
ann2$addItems(data)
# Retrieve the 4 nearest neighbors for every item in data
res2 <- ann2$getAllNNsList(data, 4, TRUE)
# labels of the data are in res$item, distances in res$distance

# If you are able to store your data column-wise, then the overhead of copying
# the data into a form usable by hnsw can be noticeably reduced
data_by_col <- t(data)
ann3 <- new(HnswL2, dim, nitems, M, ef)
ann3$addItemsCol(data_by_col)
# Retrieve the 4 nearest neighbors for every item in data_by_col
res3 <- ann3$getAllNNsListCol(data_by_col, 4, TRUE)
# The returned neared neighbor data matrices are also returned column-wise
all(res2$item == t(res3$item) & res2$distance == t(res3$distance))

# Save the index
ann$save("iris.hnsw")

# load it back in: you do need to know the dimension of the original data
ann4 <- new(HnswL2, dim, "iris.hnsw")
# new index should behave like the original
all(ann$getNNs(data[1, ], 4) == ann4$getNNs(data[1, ], 4))

# other distance classes:
# Cosine: HnswCosine
# Inner Product: HnswIP
# Euclidean: HnswEuclidean

Here's a rough equivalent of the serialization/deserialization example from the hnswlib README, but using the recently-added resizeIndex method to increase the size of the index after its initial specification, avoiding having to read from or write to disk:

library("RcppHNSW")
set.seed(12345)

dim <- 16
num_elements <- 100000

# Generate sample data
data <- matrix(stats::runif(num_elements * dim), nrow = num_elements)

# Split data into two batches
data1 <- data[1:(num_elements / 2), ]
data2 <- data[(num_elements / 2 + 1):num_elements, ]

# Create index
M <- 16
ef <- 10
# Set the initial index size to the size of the first batch
p <- new(HnswL2, dim, num_elements / 2, M, ef)

message("Adding first batch of ", nrow(data1), " elements")
p$addItems(data1)

# Query the elements for themselves and measure recall:
idx <- p$getAllNNs(data1, k = 1)
message("Recall for the first batch: ", formatC(mean(idx == 1:nrow(data1))))

# Increase the total capacity, so that it will handle the new data
p$resizeIndex(num_elements)

message("Adding the second batch of ", nrow(data2), " elements")
p$addItems(data2)

# Query the elements for themselves and measure recall:
idx <- p$getAllNNs(data, k = 1)
# You can get distances with:
# res <- p$getAllNNsList(data, k = 1, include_distances = TRUE)
# res$dist contains the distance matrix, res$item stores the indexes

message("Recall for two batches: ", formatC(mean(idx == 1:num_elements)))

Although there's no longer any need for this, for completeness, here's how you would use save and new to achieve the same effect without resizeIndex:

filename <- "first_half.bin"
# Serialize index
p$save(filename)

# Reinitialize and load the index
rm(p)
message("Loading index from ", filename)
# Increase the total capacity, so that it will handle the new data
p <- new(HnswL2, dim, filename, num_elements)
unlink(filename)

API

DO NOT USE NAMED PARAMETERS

Because these are wrappers around C++ code, you cannot use named parameters in the calling R code. Arguments are parsed by position. This is most annoying in constructors, which take multiple integer arguments, e.g.

### DO THIS ###
dim <- 10
num_elements <- 100
M <- 200
ef_construction <- 16
index <- new(HnswL2, dim, num_elements, M, ef_construction)

### DON'T DO THIS ###
index <- new(HnswL2, dim, ef_construction = 16, M = 200, num_elements = 100)
# treated as if you wrote:
index <- new(HnswL2, dim, 16, 200, 100)

OK onto the API

• new(HnswL2, dim, max_elements, M = 16, ef_contruction = 200) creates a new index using the squared L2 distance (i.e. square of the Euclidean distance), with dim dimensions and a maximum size of max_elements items. ef and M determine the speed vs accuracy trade off. Other classes for different distances are: HnswCosine for the cosine distance and HnswIp for the "Inner Product" distance (like the cosine distance without normalizing).
• new(HnswL2, dim, filename) load a previously saved index (see save below) with dim dimensions from the specified filename.
• new(HnswL2, dim, filename, max_elements) load a previously saved index (see save below) with dim dimensions from the specified filename, and a new maximum capacity of max_elements. This is a way to increase the capacity of the index without a complete rebuild.
• setEf(ef) set search parameter ef.
• setNumThreads(num_threads) Use (at most) this number of threads when adding items (via addItems) and searching the index (via getAllNNs and getAllNNsList). See also the setGrainSize parameter.
• setGrainSize(grain_size) The minimum amount of work to do (adding or searching items) per thread. If you don't have enough work for all the threads specified by setNumThreads to process grain_size items per thread, then fewer threads will be used. This is useful for cases where the cost of context switching between larger number of threads would outweigh the performance gain from parallelism. For example, if you have 100 items to process and asked for four threads, then 25 items will be processed per thread. However, setting the grain_size to 50 will result in 50 items being processed per thread, and therefore only two threads being used.
• addItem(v) add vector v to the index. Internally, each vector gets an increasing integer label, with the first vector added getting the label 1, the second 2 and so on. These labels are returned in getNNs and related methods to identify which vector in the index are neighbors.
• addItems(m) add the row vectors of the matrix m to the index. Internally, each row vector gets an increasing integer label, with the first row added getting the label 1, the second 2 and so on. These labels are returned in getNNs and related methods to identify which vector in the index are neighbors. The number of threads specified by setNumThreads is used for building the index and may be non-deterministic.
• addItemsCol(m) Like addItems but adds the column vectors of m to the index. Storing data column-wise makes copying the data for use by hnsw more efficient.
• save(filename) saves an index to the specified filename. To load an index, use the new(HnswL2, dim, filename) constructor (see above).
• getItems(ids) returns a matrix where each row is the data vector from the index associated with integer indices in the vector of ids. For cosine similarity, the l2 row-normalized vectors are returned. ids are one-indexed, i.e. to get the first and tenth vectors that were added to the index, use getItems(c(1, 10)), not getItems(c(0, 9)).
• getNNs(v, k) return a vector of the labels of the k-nearest neighbors of the vector v. Labels are integers numbered from one, representing the insertion order into the index, e.g. the label 1 represents the first item added to the index. If k neighbors can't be found, an error will be thrown. This normally means that ef or M have been set too small, but also bear in mind that you can't return more items than were put into the index.
• getNNsList(v, k, include_distances = FALSE) return a list containing a vector named item with the labels of the k-nearest neighbors of the vector v. Labels are integers numbered from one, representing the insertion order into the index, e.g. the label 1 represents the first item added to the index. If include_distances = TRUE then also return a vector distance containing the distances. If k neighbors can't be found, an error is thrown.
• getAllNNs(m, k) return a matrix of the labels of the k-nearest neighbors of each row vector in m. Labels are integers numbered from one, representing the insertion order into the index, e.g. the label 1 represents the first item added to the index. If k neighbors can't be found, an error is thrown. The number of threads specified by setNumThreads is used for searching.
• getAllNNsList(m, k, include_distances = FALSE) return a list containing a matrix named item with the labels of the k-nearest neighbors of each row vector in m. Labels are integers numbered from one, representing the insertion order into the index, e.g. the label 1 represents the first item added to the index. If include_distances = TRUE then also return a matrix distance containing the distances. If k neighbors can't be found, an error is thrown. The number of threads specified by setNumThreads is used for searching.
• getAllNNsCol(m, k) like getAllNNs but each item to be searched in m is stored by column, not row. In addition the returned matrix of k-nearest neighbors is also stored column-wise: i.e. the dimension of the return value matrix is k x n where n is the number of items (columns) in m. By passing the data column-wise, some overhead associated with copying data to and from hnsw can be reduced.
• getAllNNsListCol(m, k) like getAllNNsList but each item to be searched in m is stored by column, not row. In addition, the matrices in the returned list are also stored column-wise: i.e. the dimension of the return value matrix is k x n where n is the number of items (columns) in m. By passing the data column-wise, some overhead associated with copying data to and from hnsw can be reduced.
• size() returns the number of items in the index. This is an upper limit on the number of neighbors you can expect to return from getNNs and the other search methods.
• markDeleted(i) marks the item with label i (the ith item added to the index) as deleted. This means that the item will not be returned in any further searches of the index. It does not reduce the memory used by the index. Calls to size() do not reflect the number of marked deleted items.
• resize(max_elements) changes the maximum capacity of the index to max_elements.

Differences from Python Bindings

• Arbitrary integer labeling is not supported. Where labels are used, e.g. in the return value of getNNsList or as input in markDeleted or getItems, the labels represent the order in which the items were added to the index, using 1-indexing to be consistent with R. So in the Python bindings, the first item in the index has a default of label 0, but here it will have label 1.
• The interface roughly follows the Python one but deviates with naming and also rolls the declaration and initialization of the index into one call. And as noted above, you must pass arguments by position, not keyword.

License

GPL-3 or later.