Don't use global np.random.seed in tests #13345
Description
A few tests and doctests are currently setting np.random.seed globally. Unless I'm missing something, we should avoid that as it has the potential to both
- affect other tests (if those don't set the random seed explicitly
- hurt reproducibility: e.g. tests that don't set a
random_statemay behave deterministically because of this when run as part of the full test suite (given that test order is deterministic now) but yield another result when run separately.
sklearn/random_projection.py
591: >>> np.random.seed(42)
sklearn/covariance/empirical_covariance_.py
123: >>> np.random.seed(0)
sklearn/covariance/shrunk_covariance_.py
106: >>> np.random.seed(0)
374: >>> np.random.seed(0)
sklearn/covariance/robust_covariance.py
589: >>> np.random.seed(0)
sklearn/linear_model/huber.py
199: >>> np.random.seed(0)
sklearn/linear_model/sag.py
204: >>> np.random.seed(0)
sklearn/linear_model/ridge.py
643: >>> np.random.seed(0)
sklearn/linear_model/stochastic_gradient.py
1562: >>> np.random.seed(0)
sklearn/model_selection/_search.py
231: >>> np.random.seed(0)
sklearn/__init__.py
95: np.random.seed(_random_seed)
sklearn/svm/classes.py
877: >>> np.random.seed(0)
1004: >>> np.random.seed(0)
sklearn/neighbors/binary_tree.pxi
303: >>> np.random.seed(0)
317: >>> np.random.seed(0)
331: >>> np.random.seed(0)
344: >>> np.random.seed(1)
353: >>> np.random.seed(0)
sklearn/decomposition/tests/test_fastica.py
57: np.random.seed(0)
sklearn/neighbors/tests/test_ball_tree.py
155: np.random.seed(0)