The nmslibR package is a wrapper of NMSLIB, which according to the authors “… is a similarity search library and a toolkit for evaluation of similarity search methods. The goal of the project is to create an effective and comprehensive toolkit for searching in generic non-metric spaces. Being comprehensive is important, because no single method is likely to be sufficient in all cases. Also note that exact solutions are hardly efficient in high dimensions and/or non-metric spaces. Hence, the main focus is on approximate methods”.
I’ve searched for some time (before wrapping NMSLIB) for a nearest neighbor library which can work with high dimensional data and can scale with big datasets. I’ve already written a package for k-nearest-neighbor search (KernelKnn), however, it’s based on brute force and unfortunately, it requires a certain computation time if the data consists of many rows. The nmslibR package, besides the main functionality of the NMSLIB python library, also includes an Approximate Kernel k-nearest function, which as I will show in the next lines is both fast and accurate. A comparison of NMSLIB with other popular approximate k-nearest-neighbor methods can be found here.
The NMSLIB Library,
Details can be found in the NMSLIB-manual.
The nmslibR package includes the following R6-class / functions,
| NMSlib |
|---|
| Knn_Query() |
| knn_Query_Batch() |
| save_Index() |
UPDATE 10-05-2018 : Beginning from version 1.0.2 the dgCMatrix_2scipy_sparse function was renamed to TO_scipy_sparse and now accepts either a dgCMatrix or a dgRMatrix as input. The appropriate format for the nmslibR package in case of sparse matrices is the dgRMatrix format (scipy.sparse.csr_matrix)
| KernelKnn_nmslib() |
|---|
| KernelKnnCV_nmslib() |
|---|
| TO_scipy_sparse() |
|---|
| mat_2scipy_sparse() |
|---|
The package documentation includes details and examples for the R6-class and functions. I’ll start explaining how a user can work with sparse matrices as the input can also be a python sparse matrix.
The nmslibR package includes two functions (mat_2scipy_sparse and TO_scipy_sparse) which allow the user to convert from a matrix / sparse matrix (dgCMatrix, dgRMatrix) to a scipy sparse matrix (scipy.sparse.csc_matrix, scipy.sparse.csr_matrix),
library(nmslibR)
# conversion from a matrix object to a scipy sparse matrix
#----------------------------------------------------------
set.seed(1)
x = matrix(runif(1000), nrow = 100, ncol = 10)
x_sparse = mat_2scipy_sparse(x, format = "sparse_row_matrix")
print(dim(x))
[1] 100 10
print(x_sparse$shape)
(100, 10)# conversion from a dgCMatrix object to a scipy sparse matrix
#-------------------------------------------------------------
data = c(1, 0, 2, 0, 0, 3, 4, 5, 6)
# 'dgCMatrix' sparse matrix
#--------------------------
dgcM = Matrix::Matrix(data = data, nrow = 3,
ncol = 3, byrow = TRUE,
sparse = TRUE)
print(dim(dgcM))
[1] 3 3
x_sparse = TO_scipy_sparse(dgcM)
print(x_sparse$shape)
(3, 3)
# 'dgRMatrix' sparse matrix
#--------------------------
dgrM = as(dgcM, "RsparseMatrix")
class(dgrM)
# [1] "dgRMatrix"
# attr(,"package")
# [1] "Matrix"
print(dim(dgrM))
[1] 3 3
res_dgr = TO_scipy_sparse(dgrM)
print(res_dgr$shape)
(3, 3)The parameter settings for the NMSlib R6-class can be found in the Non-Metric Space Library (NMSLIB) Manual, which explains the NMSLIB Library in detail. In the following code chunk, I’ll show the functionality of the methods included using a data set from my Github repository (it appears as .ipynb notebook in the nmslib Github repository)
library(nmslibR)
# download the data from my Github repository (tested on a Linux OS)
#-------------------------------------------------------------------
system("wget https://raw.githubusercontent.com/mlampros/DataSets/master/sift_10k.txt")
# load the data in the R session
#-------------------------------
sift_10k = read.table("~/sift_10k.txt", quote="\"", comment.char="")
# index parameters
#-----------------
M = 15
efC = 100
num_threads = 5
index_params = list('M'= M, 'indexThreadQty' = num_threads, 'efConstruction' = efC,
'post' = 0, 'skip_optimized_index' = 1 )
# query-time parameters
#----------------------
efS = 100
query_time_params = list('efSearch' = efS)
# Number of neighbors
#--------------------
K = 100
# space to use
#---------------
space_name = 'l2sqr_sift'
# initialize NMSlib [ the data should be a matrix ]
#--------------------------------------------------
init_nms = NMSlib$new(input_data = as.matrix(sift_10k), Index_Params = index_params,
Time_Params = query_time_params, space = space_name,
space_params = NULL, method = 'hnsw',
data_type = 'DENSE_UINT8_VECTOR', dtype = 'INT',
index_filepath = NULL, print_progress = FALSE)# returns a 1-dimensional vector
#-------------------------------
init_nms$Knn_Query(query_data_row = as.matrix(sift_10k[1, ]), k = 5)[[1]]
[1] 2 6 4585 9256 140 # indices
[[2]]
[1] 18724 24320 68158 69067 70321 # distances# returns knn's for all data
#---------------------------
all_dat = init_nms$knn_Query_Batch(as.matrix(sift_10k), k = 5, num_threads = 1)
str(all_dat)# a list of indices and distances for all observations
#------------------------------------------------------
List of 2
$ knn_idx : num [1:10000, 1:5] 3 4 1 2 13 14 1 2 30 31 ...
$ knn_dist: num [1:10000, 1:5] 18724 14995 18724 14995 21038 ...Details on the various methods and parameter settings can be found in the manual of the NMSLIB python Library.
In the Vignette of the KernelKnn (Image classification of the MNIST and CIFAR-10 data using KernelKnn and HOG (histogram of oriented gradients)) package I experimented with the mnist dataset and a cross-validated kernel k-nearest-neighbors model gave 98.4 % accuracy based on HOG (histogram of oriented gradients) features. However, it took almost 30 minutes (depending on the system configuration) to complete using 6 threads. I’ve implemented a similar function using NMSLIB (KernelKnnCV_nmslib), so in the next code chunk I’ll use the same parameter setting and I’ll compare computation time and accuracy.
First load the data,
# using system('wget..') on a linux OS
system("wget https://raw.githubusercontent.com/mlampros/DataSets/master/mnist.zip")
mnist <- read.table(unz("mnist.zip", "mnist.csv"), nrows = 70000, header = T,
quote = "\"", sep = ",")X = mnist[, -ncol(mnist)]
dim(X)
## [1] 70000 784
# the 'KernelKnnCV_nmslib' function requires that the labels are numeric and start from 1 : Inf
y = mnist[, ncol(mnist)] + 1
table(y)
## y
## 1 2 3 4 5 6 7 8 9 10
## 6903 7877 6990 7141 6824 6313 6876 7293 6825 6958
# evaluation metric
acc = function (y_true, preds) {
out = table(y_true, max.col(preds, ties.method = "random"))
acc = sum(diag(out))/sum(out)
acc
}then compute the HOG features,
library(OpenImageR)
hog = HOG_apply(X, cells = 6, orientations = 9, rows = 28, columns = 28, threads = 6)
##
## time to complete : 2.101281 secs
dim(hog)
## [1] 70000 324then compute the approximate kernel k-nearest-neighbors using the cosine distance,
# parameters for 'KernelKnnCV_nmslib'
#------------------------------------
M = 30
efC = 100
num_threads = 6
index_params = list('M'= M, 'indexThreadQty' = num_threads, 'efConstruction' = efC,
'post' = 0, 'skip_optimized_index' = 1 )
efS = 100
query_time_params = list('efSearch' = efS)
# approximate kernel knn
#-----------------------
fit_hog = KernelKnnCV_nmslib(hog, y, k = 20, folds = 4, h = 1,
weights_function = 'biweight_tricube_MULT',
Levels = sort(unique(y)), Index_Params = index_params,
Time_Params = query_time_params, space = "cosinesimil",
space_params = NULL, method = "hnsw", data_type = "DENSE_VECTOR",
dtype = "FLOAT", index_filepath = NULL, print_progress = FALSE,
num_threads = 6, seed_num = 1)
# cross-validation starts ..
# |=================================================================================| 100%
# time to complete : 32.88805 secs
str(fit_hog)List of 2
$ preds:List of 4
..$ : num [1:17500, 1:10] 0 0 0 0 0 0 0 0 0 0 ...
..$ : num [1:17500, 1:10] 0 0 0 0 1 ...
..$ : num [1:17500, 1:10] 0 0 0 0 0 ...
..$ : num [1:17500, 1:10] 0 0 0 0 0 0 0 0 0 0 ...
$ folds:List of 4
..$ fold_1: int [1:17500] 49808 21991 42918 7967 49782 28979 64440 49809 30522 36673 ...
..$ fold_2: int [1:17500] 51122 9469 58021 45228 2944 58052 65074 17709 2532 31262 ...
..$ fold_3: int [1:17500] 33205 40078 68177 32620 52721 18981 19417 53922 19102 67206 ...
..$ fold_4: int [1:17500] 28267 41652 28514 34525 68534 13294 48759 47521 69395 41408 ...acc_fit_hog = unlist(lapply(1:length(fit_hog$preds),
function(x) acc(y[fit_hog$folds[[x]]],
fit_hog$preds[[x]])))
acc_fit_hog
## [1] 0.9768000 0.9786857 0.9763429 0.9760000
cat('mean accuracy for hog-features using cross-validation :', mean(acc_fit_hog), '\n')
## mean accuracy for hog-features using cross-validation : 0.9769571It took approx. 33 seconds to return with an accuracy of 97.7 % . Almost 47 times faster than KernelKnn’s corresponding function (brute force) with a slight lower accuracy rate (the braycurtis distance metric might be better suited for this dataset).
I also run the corresponding brute-force algorithm of the NMSLIB Library by setting the method parameter to seq_search,
# brute force of NMSLIB [ here we set 'Index_Params' and 'Time_Params' to NULL ]
#----------------------
fit_hog_seq = KernelKnnCV_nmslib(hog, y, k = 20, folds = 4, h = 1,
weights_function = 'biweight_tricube_MULT',
Levels = sort(unique(y)), Index_Params = NULL,
Time_Params = NULL, space = "cosinesimil",
space_params = NULL, method = "seq_search",
data_type = "DENSE_VECTOR", dtype = "FLOAT",
index_filepath = NULL, print_progress = FALSE,
num_threads = 6, seed_num = 1)
# cross-validation starts ..
# |=================================================================================| 100%
# time to complete : 4.506177 mins
acc_fit_hog_seq = unlist(lapply(1:length(fit_hog_seq$preds),
function(x) acc(y[fit_hog_seq$folds[[x]]],
fit_hog_seq$preds[[x]])))
acc_fit_hog_seq
## [1] 0.9785143 0.9802286 0.9783429 0.9784571
cat('mean accuracy for hog-features using cross-validation :', mean(acc_fit_hog_seq), '\n')
## mean accuracy for hog-features using cross-validation : 0.9788857The brute-force algorithm of the NMSLIB Library is almost 6 times faster than KernelKnn giving an accuracy of approx. 97.9 %.