Question

我的Macbook上的scikit-learn中出现了一些奇怪的随机性问题。（OS X 10.12.6，conda环境python 2.7）。作为测试，我设置了以下脚本：

import numpy.random as npr
import numpy.testing as npt
from sklearn.linear_model import LogisticRegression
from sklearn.datasets import load_breast_cancer


def test_randomness_one():
    npr.seed(54)
    rand_ints_one = npr.randint(500, size=50)
    npr.seed(54)
    rand_ints_two = npr.randint(500, size=50)
    npt.assert_array_equal(rand_ints_one, rand_ints_two)


def test_logit_one():
    data = load_breast_cancer()
    preds_one = LogisticRegression(random_state=2)\
                .fit(data['data'], data['target'])\
                .decision_function(data['data'])
    preds_two = LogisticRegression(random_state=2)\
                .fit(data['data'], data['target'])\
                .decision_function(data['data'])
    npt.assert_array_equal(preds_one, preds_two)


def test_logit_two():
    data = load_breast_cancer()
    preds_one = LogisticRegression()\
                .fit(data['data'], data['target'])\
                .decision_function(data['data'])
    preds_two = LogisticRegression()\
                .fit(data['data'], data['target'])\
                .decision_function(data['data'])
    npt.assert_array_equal(preds_one, preds_two)

# A note: main used for testing with the interpreter directly
# Executed with pytest *without* the below lines.
if __name__ == "__main__":
    test_randomness_one()
    test_logit_one()
    test_logit_two()

理论上，所有这些结果应该是相同的，运行Ubuntu和windows box的同事已经验证了这一点。在我的框中，如果在REPL中执行或通过python toy_test.py运行，则所有这些测试都会通过。但是，如果通过pytest toy_test.py运行，则test_logit_one会一直失败，而test_logit_two会经常失败，但并非总是如此。在这种情况下，随机性来自哪里？它是操作系统级别的吗？ conda - 水平？ pytest？或其他什么？

为什么py.test忽略os x上的随机播种？

0 个答案: