oll 0.2.1

Online binary classification algorithms library (wrapper for OLL C++ library)

oll-python

This is a Python binding of the OLL library for machine learning.

Currently, OLL 0.03 supports following binary classification algorithms:

• Perceptron

• Averaged Perceptron

• Passive Agressive (PA, PA-I, PA-II)

• ALMA (modified slightly from original)

• Confidence Weighted Linear-Classification.

For details of oll, see: http://code.google.com/p/oll

Installation

$ pip install oll

OLL library is bundled, so you don’t need to install it separately.

Usage

import oll
# You can choose algorithms in
# "P" -> Perceptron,
# "AP" -> Averaged Perceptron,
# "PA" -> Passive Agressive,
# "PA1" -> Passive Agressive-I,
# "PA2" -> Passive Agressive-II,
# "PAK" -> Kernelized Passive Agressive,
# "CW" -> Confidence Weighted Linear-Classification,
# "AL" -> ALMA
o = oll.oll("CW", C=1.0, bias=0.0)
o.add({0: 1.0, 1: 2.0, 2: -1.0}, 1)  # train
o.classify({0:1.0, 1:1.0})  # predict
o.save('oll.model')
o.load('oll.model')

# scikit-learn like fit/predict interface
import numpy as np
array = np.array([[1, 2, -1], [0, 0, 1]])
o.fit(array, [1, -1])
o.predict(np.array([[1, 2, -1], [0, 0, 1]]))
# => [1, -1]
from scipy.sparse import csr_matrix
matrix = csr_matrix([[1, 2, -1], [0, 0, 1]])
o.fit(matrix, [1, -1])
o.predict(matrix)
# => [1, -1]

# Multi label classification
import time
import oll
from sklearn.multiclass import OutputCodeClassifier
from sklearn import datasets, cross_validation, metrics


dataset = datasets.load_digits()
ALGORITHMS = ("P", "AP", "PA", "PA1", "PA2", "PAK", "CW", "AL")
for algorithm in ALGORITHMS:
    print(algorithm)
    occ_predicts = []
    expected = []
    start = time.time()
    for (train_idx, test_idx) in cross_validation.StratifiedKFold(dataset.target,
                                                                  n_folds=10, shuffle=True):
        clf = OutputCodeClassifier(oll.oll(algorithm))
        clf.fit(dataset.data[train_idx], dataset.target[train_idx])
        occ_predicts += list(clf.predict(dataset.data[test_idx]))
        expected += list(dataset.target[test_idx])
    print('Elapsed time: %s' % (time.time() - start))
    print('Accuracy', metrics.accuracy_score(expected, occ_predicts))
# => P
# => Elapsed time: 109.82188701629639
# => Accuracy 0.770172509738
# => AP
# => Elapsed time: 111.42936396598816
# => Accuracy 0.760155815248
# => PA
# => Elapsed time: 110.95964503288269
# => Accuracy 0.74735670562
# => PA1
# => Elapsed time: 111.39844799041748
# => Accuracy 0.806343906511
# => PA2
# => Elapsed time: 115.12716913223267
# => Accuracy 0.766277128548
# => PAK
# => Elapsed time: 119.53838682174683
# => Accuracy 0.77796327212
# => CW
# => Elapsed time: 121.20785689353943
# => Accuracy 0.771285475793
# => AL
# => Elapsed time: 116.52497220039368
# => Accuracy 0.785754034502

Note

• This module requires C++ compiler to build.

• oll.cpp & oll.hpp : Copyright (c) 2011, Daisuke Okanohara

• oll_swig_wrap.cxx is generated based on ‘oll_swig.i’ in oll-ruby (https://github.com/syou6162/oll-ruby)

License

New BSD License.

CHANGES

0.2.1 (2017-6-30)

• Multi label clasification (using scikit-learn)

• Support Python 3.6

0.2 (2016-11-26)

• scikit-learn like fit/predict interfaces are available

• Setting C and bias parameters is available in initialization

• Support Python 3.5

• Unsupport Python 2.6 and 3.3

0.1.2 (2015-01-11)

• Support testFile method

• docstrings are available

0.1.1 (2014-03-29)

• Compatibility some compilers

0.1 (2013-10-11)

• Initial release.