Skip to main content

Implementation of the KSU compression algorithm

Project description

KSU Compression Algorithm Implementation

Algortihm 1 from Nearest-Neighbor Sample Compression: Efficiency, Consistency, Infinite Dimensions


  • With pip: pip install ksu
  • From source:
    • git clone --depth=1
    • cd ksu_classifier
    • python install


Command Line

This package provides two command line tools: e-net and ksu:

  • e-net constructs an epsilon net for a given epsilon
  • ksu runs the full algorithm

Both provide the -h flag to specify the arguments, and both can save the result to the disk in numpy's .npz format


This package provides a class KSU(Xs, Ys, metric, [gram, prune, logLevel, n_jobs])

Xs and Ys are the data points and their respective labels as numpy arrays

metric is either a callable to compute the metric or a string that names one of our provided metrics (print KSU.METRICS.keys() for the full list)

gram (optional, default=None) a precomputed gramian matrix, will be calculated if not provided.

prune (optional, default=False) a boolean indicating whether to prune the compressed set or not (Algorithm 2 from Near-optimal sample compression for nearest neighbors)

logLevel (optional, default='CRITICAL') a string indicating the logging level (set to 'INFO' or 'DEBUG' to get more information)

n_jobs (optional, default=1) an integer defining how many cpus to use (scipy logic), pass -1 to use all. For n_jobs below -1, (n_cpus + 1 + n_jobs) are used. Thus for n_jobs = -2, all CPUs but one are used.

KSU provides a method compressData([delta, minCompress, maxCompress, greedy, stride, logLevel, numProcs])

Which selects the subset with the lowest estimated error with confidence 1 - delta. Can take arguments:

delta (optional, default=0.1) confidence for error upper bound

minCompress (optional, default=0.05) minimal compression ratio

maxCompress (optional, default=0.1) maximum compression ratio

greedy (optional, default=True) whether to use greedy or hierarichal strategy for net construction

stride (optional, default=200) how many gammas to skip between each iteration (since similar gammas will produce similar nets)

logLevel (optional, default='CRITICAL') a string indicating the logging level (set to 'INFO' or 'DEBUG' to get more information)

numProcs (optional, default=1) number of processes to use

You can then run getClassifier() which returns a 1-NN Classifer (based on sklearn's K-NN) fitted to the compressed data.

Or, run getCompressedSet() to get the compressed data as a tuple of numpy arrays (compressedXs, compressedYs).

See scripts/ for example usage

Built-in metrics

['chebyshev', 'yule', 'sokalmichener', 'canberra', 'EarthMover', 'rogerstanimoto', 'matching', 'dice', 'EditDistance', 'braycurtis', 'russellrao', 'cosine', 'cityblock', 'l1', 'manhattan', 'sqeuclidean', 'jaccard', 'seuclidean', 'sokalsneath', 'kulsinski', 'minkowski', 'mahalanobis', 'euclidean', 'l2', 'hamming', 'correlation', 'wminkowski']

Project details

Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Files for ksu, version 0.5.1
Filename, size File type Python version Upload date Hashes
Filename, size ksu-0.5.1-py2.py3-none-any.whl (18.6 kB) File type Wheel Python version py2.py3 Upload date Hashes View
Filename, size ksu-0.5.1.tar.gz (15.4 kB) File type Source Python version None Upload date Hashes View

Supported by

Pingdom Pingdom Monitoring Google Google Object Storage and Download Analytics Sentry Sentry Error logging AWS AWS Cloud computing DataDog DataDog Monitoring Fastly Fastly CDN DigiCert DigiCert EV certificate StatusPage StatusPage Status page