svmloader

Fast loader for libsvm / svmlight sparse data files in Python (Cython), an alternative to sklearn.datasets.load_svmlight_file. svmloader loads them as scipy.sparse.csr_matrix matrices and numpy label arrays.

It is not fully equivalent to sklearn.datasets.load_svmlight_file, in particular query_id are not supported and dtype is restricted.

The types of data and labels are distinguished. The labels types supported are int and float (default int), and data can be parsed as numpy.float64 or numpy.float32 type (float64 by default).

Compressed data in .gz or .bz2 format is supported as well (see Benchmarks for the cost of bz2 decompression).

Install

Simply use pip install svmloader (Python 3.9+).

Alternatively, you can clone the repository and run pip install ..

Dependencies :

• numpy
• scipy

Usage

from svmloader import load_svmfile, load_svmfiles

# X is a scipy.sparse.csr_matrix, y a numpy array of labels.
# Column indices are zero-based by default: pass zero_based=False
# for one-based files (as in the libsvm datasets).
X, y = load_svmfile('train.libsvm', zero_based=False)

# float32 data, float labels, bz2 file decompressed on the fly
X, y = load_svmfile('train.libsvm.bz2', dtype='f', ltype='d', zero_based=False)

# several files at once, with the same number of columns
X_train, y_train, X_test, y_test = load_svmfiles(['train.libsvm', 'test.libsvm.gz'], zero_based=False)

Documentation

See here.

Benchmarks

Benchmarks on data taken from libsvm datasets, with Python 3.12, numpy 2.5, scipy 1.18 and scikit-learn 1.9.1 on an Intel Core i7-3770. Times are the best of 3 runs (2 for rcv1) in a fresh process, with the files in the page cache.

svmloader is 5-6x faster than scikit-learn on mnist and about 3x faster on rcv1. For reference, libsvm's svm_read_problem takes 5.0s on mnist.

Compressed files are decompressed on the fly. bzip2 decompression is CPU-bound and much slower than parsing: on the same machine, loading mnist.bz2 takes 3-8s (versus 0.9s uncompressed) and the 815 MB rcv1 file takes 60-100s from .bz2 (versus 7.6s uncompressed). The range reflects CPU throttling under sustained decompression; .gz files decompress much faster.