v0.4.3

Dependencies

• PyCOMPSs == 2.5
• Scikit-learn >= 0.19.2
• NumPy >= 1.15.4
• Scipy >= 1.0.0

Improvements

• Installing dislib via pip now automatically places the dislib executable in the PATH.

v0.4.0

Dependencies

• PyCOMPSs == 2.5
• Scikit-learn >= 0.19.2
• NumPy >= 1.15.4
• Scipy >= 1.0.0

Breaking Changes

• Most estimator methods, such as fit and predict, now expect one or two ds-arrays instead of a Dataset.

New Features

• This release introduces the distributed array as the main data structure in dislib. All estimators have been modified to accept ds-arrays instead of Datasets. The Dataset and Subset classes have been removed.

Bug Fixes

• Minor bug fixes in RandomForestClassifier and K-means

Improvements

• The performance of various algorithms has been improved by using PyCOMPSs COLLECTIONS.
• K-means now accepts an 'init' parameter.

v0.3.0

Dependencies

• PyCOMPSs == 2.5
• Scikit-learn >= 0.19.2
• NumPy >= 1.15.4
• Scipy >= 1.0.0

New Features

• GaussianMixture now supports covariance types 'tied', 'diag', and 'spherical' apart from 'full'.
• dislib now provides PCA and LinearRegression models.

Bug Fixes

• Fixed DBSCAN to be able to detect clusters with less than min_samples samples, and to be able to detect clusters that lie in the intersection of two regions.

Improvements

• The GaussianMixture documentation has been improved.
• Extra tests for GaussianMixture, C-SVM and DBSCAN have been added.
• The performance of K-means, DBSCAN and GaussianMixtures has been significantly improved.
• The performance of utils.shuffle has been improved by using PyCOMPSs collections.
• The performance of Dataset has been improved by removing the tracking of duplicates.

v0.2.1

Dependencies

(Update dependency versions if required)

• PyCOMPSs >= 2.4-rc1902
• Scikit-learn >= 0.19.1
• NumPy >= 1.15.4
• Scipy >= 1.0.0

Bug Fixes

• DBSCAN now detects clusters with less than min_samples in certain situations

Improvements

• The performance of DBSCAN has been improved

v0.2.0

Dependencies

• PyCOMPSs == 2.4-rc1902
• Scikit-learn => 0.19.1
• NumPy => 1.15.4
• Scipy => 1.0.0

Upgrade Steps

Breaking Changes

• predict and fit_predict methods in K-means, DBSCAN and C-SVM now take a Dataset as argument and do not return anything

New Features

• The following new algorithms have been implemented:

  • Gaussian mixtures
  • Nearest neighbors
  • Alternating least squares
  • Standard scaler

• Added the following utility methods:

  • resample
  • shuffle
  • as_grid

Bug Fixes

• Numerous bug fixes in DBSCAN.
• Fixed the reproducibility of results in C-SVM and random forests
• Several other minor bug fixes

Improvements

• Completely unified the interface of the different algorithms
• Improved the documentation
• Added a way to easily access Dataset samples and labels
• Implemented Dataset's transpose
• Implemented Dataset's apply function

0.1.1

This release has been tested with COMPSs version rc1902.

Initial Release

Merge pull request #45 from bsc-wdc/kmeans-fix

Kmeans fix