PyML

PyML is a Python package with machine learning algorithms written in Python and C/C++.

Installation

From source

git clone [email protected]:gf712/PyML.git

cd to/folder/of/pyml

python setup.py install

To verify installation you should run:

python setup.py test

From anaconda

conda install -c gf712 pyml

Usage

>>> from pyml.linear_models import LinearRegression
>>> from pyml.datasets import regression
>>> X, y = regression(seed=1970)
>>> lr = LinearRegression(solver='OLS', bias=True)
>>> lr.train(X, y)
>>> lr.coefficients
[0.3011617891659273, 0.9428803588636959]

Changelog

Version 0.2.1.1 (14/12/2017):

• Major:

  • Library now available on anaconda

• Minor:

  • Backend speed improvements

Version 0.2.1 (05/12/2017):

• Major:

  • Machine learning algorithms:

    • PCA implementation
    • Multiclass LogisticRegression

  • Optimisation algorithms:

    • Mini batch gradient descent
    • Gradient descent with momentum
    • Nesterov optimisation
    • Adagrad optimisation
    • Adadelta optimisation

  • Backend:

    • Major changes to flatArrays and C++ backend leading to speed improvements

• Minor:

  • Eigendecomposition of symmetric matrices with Jacobi rotations
  • Improved docs
  • Fixed memory leaks
  • increased coverage

Version 0.2 (16/11/2017):

• Major:

  • KMeans implementation
  • Logistic regression implementation
  • New C++ class to represent 2D arrays more efficiently

• Minor:

  • Cleaned up C++ code
  • Norm is calculated with pure C++ (Python only used to provide euclidean_distance and manhattan_distance interface)
  • Faster implementation of quick sort algorithm (with C++)
  • Fixed argsort behaviour
  • Overall speed improvements using flat arrays (e.g. matrix matrix multiplication ~2 times faster)
  • More tests and increased code coverage
  • Started fixing memory leaks
  • column and row wise sort and argsort

Version 0.1 (09/11/2017):

• Major:

  • Linear regression with ordinary least squares
  • C++ backend

• Minor:

  • Linear regression with gradient descent about 10x faster with C++ backend
  • Linear regression with OLS (in C++) several orders of magnitude faster than Python gradient descent (depending on set size)
  • KNN uses C++ for distance calculations

Version 0.1dev (02/11/2017):

• Major:
  • Linear regression with gradient descent
  • KNN, both regressor and classifier
  • Pure Python implementations