scikit-learn/sklearn/utils/tests/test_random.py

import numpy as np
import pytest
import scipy.sparse as sp
from scipy.special import comb
from numpy.testing import assert_array_almost_equal

from sklearn.utils.random import _random_choice_csc, sample_without_replacement
from sklearn.utils._random import _our_rand_r_py


###############################################################################
# test custom sampling without replacement algorithm
###############################################################################
def test_invalid_sample_without_replacement_algorithm():
    with pytest.raises(ValueError):
        sample_without_replacement(5, 4, "unknown")


def test_sample_without_replacement_algorithms():
    methods = ("auto", "tracking_selection", "reservoir_sampling", "pool")

    for m in methods:

        def sample_without_replacement_method(
            n_population, n_samples, random_state=None
        ):
            return sample_without_replacement(
                n_population, n_samples, method=m, random_state=random_state
            )

        check_edge_case_of_sample_int(sample_without_replacement_method)
        check_sample_int(sample_without_replacement_method)
        check_sample_int_distribution(sample_without_replacement_method)


def check_edge_case_of_sample_int(sample_without_replacement):

    # n_population < n_sample
    with pytest.raises(ValueError):
        sample_without_replacement(0, 1)
    with pytest.raises(ValueError):
        sample_without_replacement(1, 2)

    # n_population == n_samples
    assert sample_without_replacement(0, 0).shape == (0,)

    assert sample_without_replacement(1, 1).shape == (1,)

    # n_population >= n_samples
    assert sample_without_replacement(5, 0).shape == (0,)
    assert sample_without_replacement(5, 1).shape == (1,)

    # n_population < 0 or n_samples < 0
    with pytest.raises(ValueError):
        sample_without_replacement(-1, 5)
    with pytest.raises(ValueError):
        sample_without_replacement(5, -1)


def check_sample_int(sample_without_replacement):
    # This test is heavily inspired from test_random.py of python-core.
    #
    # For the entire allowable range of 0 <= k <= N, validate that
    # the sample is of the correct length and contains only unique items
    n_population = 100

    for n_samples in range(n_population + 1):
        s = sample_without_replacement(n_population, n_samples)
        assert len(s) == n_samples
        unique = np.unique(s)
        assert np.size(unique) == n_samples
        assert np.all(unique < n_population)

    # test edge case n_population == n_samples == 0
    assert np.size(sample_without_replacement(0, 0)) == 0


def check_sample_int_distribution(sample_without_replacement):
    # This test is heavily inspired from test_random.py of python-core.
    #
    # For the entire allowable range of 0 <= k <= N, validate that
    # sample generates all possible permutations
    n_population = 10

    # a large number of trials prevents false negatives without slowing normal
    # case
    n_trials = 10000

    for n_samples in range(n_population):
        # Counting the number of combinations is not as good as counting the
        # the number of permutations. However, it works with sampling algorithm
        # that does not provide a random permutation of the subset of integer.
        n_expected = comb(n_population, n_samples, exact=True)

        output = {}
        for i in range(n_trials):
            output[
                frozenset(sample_without_replacement(n_population, n_samples))
            ] = None

            if len(output) == n_expected:
                break
        else:
            raise AssertionError(
                "number of combinations != number of expected (%s != %s)"
                % (len(output), n_expected)
            )


def test_random_choice_csc(n_samples=10000, random_state=24):
    # Explicit class probabilities
    classes = [np.array([0, 1]), np.array([0, 1, 2])]
    class_probabilities = [np.array([0.5, 0.5]), np.array([0.6, 0.1, 0.3])]

    got = _random_choice_csc(n_samples, classes, class_probabilities, random_state)
    assert sp.issparse(got)

    for k in range(len(classes)):
        p = np.bincount(got.getcol(k).toarray().ravel()) / float(n_samples)
        assert_array_almost_equal(class_probabilities[k], p, decimal=1)

    # Implicit class probabilities
    classes = [[0, 1], [1, 2]]  # test for array-like support
    class_probabilities = [np.array([0.5, 0.5]), np.array([0, 1 / 2, 1 / 2])]

    got = _random_choice_csc(
        n_samples=n_samples, classes=classes, random_state=random_state
    )
    assert sp.issparse(got)

    for k in range(len(classes)):
        p = np.bincount(got.getcol(k).toarray().ravel()) / float(n_samples)
        assert_array_almost_equal(class_probabilities[k], p, decimal=1)

    # Edge case probabilities 1.0 and 0.0
    classes = [np.array([0, 1]), np.array([0, 1, 2])]
    class_probabilities = [np.array([0.0, 1.0]), np.array([0.0, 1.0, 0.0])]

    got = _random_choice_csc(n_samples, classes, class_probabilities, random_state)
    assert sp.issparse(got)

    for k in range(len(classes)):
        p = (
            np.bincount(
                got.getcol(k).toarray().ravel(), minlength=len(class_probabilities[k])
            )
            / n_samples
        )
        assert_array_almost_equal(class_probabilities[k], p, decimal=1)

    # One class target data
    classes = [[1], [0]]  # test for array-like support
    class_probabilities = [np.array([0.0, 1.0]), np.array([1.0])]

    got = _random_choice_csc(
        n_samples=n_samples, classes=classes, random_state=random_state
    )
    assert sp.issparse(got)

    for k in range(len(classes)):
        p = np.bincount(got.getcol(k).toarray().ravel()) / n_samples
        assert_array_almost_equal(class_probabilities[k], p, decimal=1)


def test_random_choice_csc_errors():
    # the length of an array in classes and class_probabilities is mismatched
    classes = [np.array([0, 1]), np.array([0, 1, 2, 3])]
    class_probabilities = [np.array([0.5, 0.5]), np.array([0.6, 0.1, 0.3])]
    with pytest.raises(ValueError):
        _random_choice_csc(4, classes, class_probabilities, 1)

    # the class dtype is not supported
    classes = [np.array(["a", "1"]), np.array(["z", "1", "2"])]
    class_probabilities = [np.array([0.5, 0.5]), np.array([0.6, 0.1, 0.3])]
    with pytest.raises(ValueError):
        _random_choice_csc(4, classes, class_probabilities, 1)

    # the class dtype is not supported
    classes = [np.array([4.2, 0.1]), np.array([0.1, 0.2, 9.4])]
    class_probabilities = [np.array([0.5, 0.5]), np.array([0.6, 0.1, 0.3])]
    with pytest.raises(ValueError):
        _random_choice_csc(4, classes, class_probabilities, 1)

    # Given probabilities don't sum to 1
    classes = [np.array([0, 1]), np.array([0, 1, 2])]
    class_probabilities = [np.array([0.5, 0.6]), np.array([0.6, 0.1, 0.3])]
    with pytest.raises(ValueError):
        _random_choice_csc(4, classes, class_probabilities, 1)


def test_our_rand_r():
    assert 131541053 == _our_rand_r_py(1273642419)
    assert 270369 == _our_rand_r_py(0)