def test_subset_raises_on_float():
"""[Base] SubsetIndex: check raises error on floats for n_part, n_split."""
with np.testing.assert_raises(ValueError):
SubsetIndex(0.5, 2, X=X)
with np.testing.assert_raises(ValueError):
SubsetIndex(2, 0.5, X=X)
def test_subset_partition():
"""[Base] Subset: test partition indexing on tuples."""
parts = list()
for part in SubsetIndex(X=X).partition():
parts.append(part)
assert parts == [(0, 3), (3, 5)]
def test_subset_tuple_shape():
"""[Base] SubsetIndex: test the tuple shape on generation."""
tup = [(tri, tei) for tri, tei in SubsetIndex(2, 2).generate(X)]
assert tup == [(((2, 3),), [(0, 2), (3, 4)]),
(((0, 2),), [(2, 3), (4, 5)]),
(((4, 5),), [(0, 2), (3, 4)]),
(((3, 4),), [(2, 3), (4, 5)])]
def test_subset_partition_array():
"""[Base] Subset: test partition indexing on arrays."""
parts = list()
for part in SubsetIndex(X=X).partition(as_array=True):
parts.append(part)
np.testing.assert_array_equal(parts[0], np.array([0, 1, 2]))
np.testing.assert_array_equal(parts[1], np.array([3, 4]))
def test_subset_index_is_fitted():
"""[Base] BlendIndex: check fit methods."""
attrs = ['n_samples', 'n_test_samples']
idx = SubsetIndex()
for attr in attrs: assert not getattr(idx, attr)
idx.fit(X)
for attr in attrs: assert getattr(idx, attr)
idx = SubsetIndex()
for attr in attrs: assert not getattr(idx, attr)
for _ in idx.generate(X): pass
for attr in attrs: assert getattr(idx, attr)
idx = SubsetIndex(X=X)
for attr in attrs: assert getattr(idx, attr)
def test_subset_array_shape():
"""[Base] ClusteredSubsetIndex: test the array shape on generation."""
t = list()
e = list()
for tri, tei in SubsetIndex(2, 2, X=X).generate(as_array=True):
t.append(tri.tolist())
e.append(tei.tolist())
assert t == [[2], [0, 1], [4], [3]]
assert e == [[0, 1, 3], [2, 4], [0, 1, 3], [2, 4]]
# We can create several other types of learners by
# varying the estimation strategy. An especially interesting strategy is to
# partition the training set and create several learners fitted on a given
# partition. This will create one prediction feature per partition.
# In the following example we fit the OLS model using two partitions and
# three fold CV on each partition. Note that by passing the output array
# as an argument during ``'fit'``, we perform a fit and transform operation.
from mlens.index import SubsetIndex
def mse(y, p):
return np.mean((y - p)**2)
indexer = SubsetIndex(partitions=2, folds=2, X=X)
learner = Learner(estimator=OLS(),
indexer=indexer,
name='subsemble-ols',
scorer=mse,
verbose=True)
job.job = 'fit'
job.predict_out = np.zeros((y.shape[0], 2))
learner.setup(job.predict_in, job.targets, job.job)
for sub_learner in learner(job.args(), 'main'):
sub_learner.fit()
print('Output:')
print(job.predict_out)
print()
def test_subset_raises_empty():
"""[Base] SubsetIndex: check raises error on empty train set."""
with np.testing.assert_raises(ValueError):
SubsetIndex(2, 2, X=np.empty(1))
def test_subset_raises_no_partition():
"""[Base] SubsetIndex: check raises error on 0 partitions."""
with np.testing.assert_raises(ValueError):
SubsetIndex(0, X=X)
def test_subset_raises_on_no_split_part():
"""[Base] SubsetIndex: check raises error n_part * n_split > n_samples."""
with np.testing.assert_raises(ValueError):
SubsetIndex(3, 3, X=X)
def test_subset_raises_opartitions_and_one_split():
"""[Base] SubsetIndex: check raises error on single split of partitions."""
with np.testing.assert_raises(ValueError):
SubsetIndex(2, 1, X=X)
def test_subset_raises_on_w_raise_():
"""[Base] SubsetIndex: check raises on n_part = 1, folds = 1."""
with np.testing.assert_raises(ValueError):
SubsetIndex(1, 1, X=X)
def test_subset_warns_on_wo_raise_():
"""[Base] SubsetIndex: check raises on n_part = 1, folds = 1."""
with np.testing.assert_warns(UserWarning):
SubsetIndex(1, 1, raise_on_exception=False, X=X)
