Efficient library for reading and writing labelled sparse matrices in delimited text format (a.k.a. SVMLight format), as used by software such as SVMLight, LibSVM, ThunderSVM, LibFM, xLearn, XGBoost, LightGBM, and others. Supports labels for regression, classification (binary, multi-class, and multi-label), and ranking (with qid field).

Written in C++ with interfaces for Python and R.

Timings for reading the Cover Type dataset, available at LibSVM datasets.

The aim of the library is to read and write sparse CSR matrices into text in a delimited format (which is used by popular machine learning software) in which each row is represented by a line, with the data written as follows:

<label(s)> <column>:<value> <column>:<value> ...

Example line (row):

1 1:1.234 3:20

This line denotes a row with label (target variable) equal to 1, a value for the first column of 1.234, a value of zero for the second column (which is missing), and a value of 20 for the third column.

Lines can come in slightly different formats according to the desired task:

• Regression

0.321 2:1.21 5:2.05
-1.234 1:0.45 3:0.001 4:-10

• Classification

1 2:1.21 5:2.05
-1 1:0.45 3:0.001 4:-10

• Multi-Label classification

1,2,3 2:1.21 5:2.05
2 1:0.45 3:0.001 4:-10

• Ranking

1 qid:1 2:1.21 5:2.05
2 qid:2 1:0.45 3:0.001 4:-10

Lines might also contain comments (everything after a # is considered a comment) and/or a header with metadata (number of rows, columns, and classes).

• Python

Note: requires C/C++ compilers configured for Python. See this guide for instructions.

pip install readsparse

or if that fails:

pip install --no-use-pep517 readsparse

(A small note: on Windows, if compiling with MinGW, will use its default stdio library, which at the time of writing takes it from an outdated MSVC library. To use MinGW's own workarounds for stdio, one can define an environment variable ANSISTDIO or pass argument -ansistdio to setup.py)

IMPORTANT: the setup script will try to add compilation flag -march=native. This instructs the compiler to tune the package for the CPU in which it is being installed (by e.g. using AVX instructions if available), but the result might not be usable in other computers. If building a binary wheel of this package or putting it into a docker image which will be used in different machines, this can be overriden either by (a) defining an environment variable DONT_SET_MARCH=1, or by (b) manually supplying compilation CFLAGS as an environment variable with something related to architecture. For maximum compatibility (but slowest speed), it's possible to do something like this:

export DONT_SET_MARCH=1
pip install readsparse

or, by specifying some compilation flag for architecture:

export CFLAGS="-march=x86-64"
export CXXFLAGS="-march=x86-64"
pip install readsparse

• R

install.packages("readsparse")

• C++

git clone https://www.github.com/david-cortes/readsparse.git
mkdir build
cd build
cmake -DUSE_MARCH_NATIVE=1 ..
cmake --build .
sudo make install

(If adding -DUSE_MARCH_NATIVE=1, will add argument -march=native, which will optimize it for the CPU on which it's being compiled and might not run on older CPUs. Remove that option for more "portability")

• Python: documentation is available at ReadTheDocs.

• R: documentation is available at CRAN.

• C++: documentation is available under the public header.

• Python

import numpy as np
import readsparse
import scipy.sparse as sp

coded_matrix = """
1 2:1.21 5:2.05
-1 1:0.45 3:0.001 4:-10
""".strip()

r = readsparse.read_sparse(coded_matrix, from_string=True)
r

{'X': <2x5 sparse matrix of type '<class 'numpy.float64'>'
    with 5 stored elements in Compressed Sparse Row format>,
 'y': array([ 1., -1.])}

r["X"].toarray()

array([[ 0.00e+00,  1.21e+00,  0.00e+00,  0.00e+00,  2.05e+00],
       [ 4.50e-01,  0.00e+00,  1.00e-03, -1.00e+01,  0.00e+00]])

### Convert it back to text
recoded_martix = readsparse.write_sparse(file=None, X=r["X"], y=r["y"], to_string=True)
print(recoded_martix)

1.00000000 2:1.21000000 5:2.05000000
-1.00000000 1:0.45000000 3:0.00100000 4:-10.00000000

### Example with file I/O
## generate a random sparse matrix and labels
np.random.seed(1)
X = sp.random(m=5, n=10, density=0.2)
y = np.random.normal(size=5)

## save into a text file
temp_file = "matrix.txt"
readsparse.write_sparse(temp_file, X, y, integer_labels=False)

## inspect the text file
with open(temp_file, "r") as f:
    print(f.read())

-1.27321995 8:0.67165410 9:0.41178788
1.01498680 7:0.62402999 10:0.03417131
-1.48105971 1:0.28962964 6:0.59306552
-0.28709989 1:0.14212014 8:0.19755090 10:0.78331447
-0.05682428 6:0.41253884

## read it back
r = readsparse.read_sparse(temp_file)
r

{'X': <5x10 sparse matrix of type '<class 'numpy.float64'>'
    with 10 stored elements in Compressed Sparse Row format>,
 'y': array([-1.27321995,  1.0149868 , -1.48105971, -0.28709989, -0.05682428])}

• R

(Example also available under the internal documentation)

library(Matrix)
library(readsparse)

### Example input file
"1 2:1.21 5:2.05
-1 1:0.45 3:0.001 4:-10" -> coded.matrix

r <- read.sparse(coded.matrix, from_string=TRUE)
print(r)

$X
2 x 5 sparse Matrix of class "dgRMatrix"
                             
[1,] .    1.21 .       . 2.05
[2,] 0.45 .    0.001 -10 .   

$y
[1]  1 -1

### Convert it back to text
recoded.matrix <- write.sparse(file=NULL, X=r$X, y=r$y, to_string=TRUE)
cat(recoded.matrix)

1 2:1.21000000 5:2.05000000
-1 1:0.45000000 3:0.00100000 4:-10.00000000

### Example with file I/O
## generate a random sparse matrix and labels
set.seed(1)
X <- rsparsematrix(nrow=5, ncol=10, nnz=8)
y <- rnorm(5)

## save into a text file
temp_file <- file.path(tempdir(), "matrix.txt")
write.sparse(temp_file, X, y, integer_labels=FALSE)

## inspect the text file
cat(paste(readLines(temp_file), collapse="\n"))

-2.21469989 1:0.74000000
1.12493092 
-0.04493361 4:-0.62000000 5:-0.31000000 9:1.50000000
-0.01619026 1:-0.82000000 3:0.39000000 7:0.58000000 8:0.49000000
0.94383621

## read it back
r <- read.sparse(temp_file)
print(r)

$X
5 x 9 sparse Matrix of class "dgRMatrix"
                                             
[1,]  0.74 . .     .     .    . .    .    .  
[2,]  .    . .     .     .    . .    .    .  
[3,]  .    . .    -0.62 -0.31 . .    .    1.5
[4,] -0.82 . 0.39  .     .    . 0.58 0.49 .  
[5,]  .    . .     .     .    . .    .    .  

$y
[1] -2.21469989  1.12493092 -0.04493361 -0.01619026  0.94383621

• C++

See file readsparse_cpp_example.cpp

Public datasets in this format can be found under the LibSVM datasets and the Extreme Classification Repository.