BUG: ShuffleSplit should give reproducible splits

when seeded with a controled seed.

Author: GaelVaroquaux

sklearn/cross_validation.py

@@ -632,7 +632,7 @@ def __init__(self, n, n_bootstraps=3, n_train=0.5, n_test=None,
  self.random_state = random_state
 
  def __iter__(self):
- rng = self.random_state = check_random_state(self.random_state)
+ rng = check_random_state(self.random_state)
  for i in range(self.n_bootstraps):
  # random partition
  permutation = rng.permutation(self.n)
@@ -743,7 +743,7 @@ def __init__(self, n, n_iterations=10, test_fraction=0.1,
  (train_fraction, test_fraction))
 
  def __iter__(self):
- rng = self.random_state = check_random_state(self.random_state)
+ rng = check_random_state(self.random_state)
  n_test = ceil(self.test_fraction * self.n)
  if self.train_fraction is None:
  n_train = self.n - n_test

sklearn/tests/test_cross_validation.py

@@ -5,6 +5,7 @@
 
 from nose.tools import assert_true
 from nose.tools import assert_raises
+from nose.tools import assert_equal
 
 from ..base import BaseEstimator
 from ..datasets import make_regression
@@ -230,6 +231,13 @@ def test_shufflesplit_errors():
  test_fraction=0.1, train_fraction=0.95)
 
 
+def test_shufflesplit_reproducible():
+ # Check that iterating twice on the ShuffleSplit gives the same
+ # sequence of train-test when the random_state is given
+ ss = cross_validation.ShuffleSplit(10, random_state=21)
+ assert_array_equal(list(a for a, b in ss), list(a for a, b in ss))
+
+
 def test_cross_indices_exception():
  X = coo_matrix(np.array([[1, 2], [3, 4], [5, 6], [7, 8]]))
  y = np.array([1, 1, 2, 2])