def test_not_constatnt_splitter():
sequence = np.array([1, 1, 1, 2, 2, 2, 2])
splitter = pasio.NotZeroSplitter(base_splitter=pasio.SquareSplitter())
splits = splitter.split(sequence, simple_scorer_factory)
assert splits[1] == [0, 3]
splitter = pasio.NotZeroSplitter(base_splitter=pasio.SquareSplitter())
splits = splitter.split(sequence, simple_greedy_scorer_factory)
assert splits[1] == list(range(len(sequence)))
splitter = pasio.NotConstantSplitter(base_splitter=pasio.SquareSplitter())
splits = splitter.split(sequence, simple_greedy_scorer_factory)
assert splits[1] == [0, 3]
splitter = pasio.NotConstantSplitter(base_splitter=pasio.SquareSplitter())
splits = splitter.split(sequence, simple_greedy_scorer_factory,
np.array(range(len(sequence) - 1)))
assert splits[1] == [0, 3]
splitter = pasio.NotConstantSplitter(base_splitter=pasio.SquareSplitter())
assert np.allclose(
np.array([0, 3]),
splitter.get_non_constant_split_candidates(sequence, None))
assert np.allclose(
np.array([0, 3]),
splitter.get_non_constant_split_candidates(sequence, np.array([0, 3])))
def test_split_with_length_regularization():
# score of split 'AAA|B|AA' = 9+1+4 = 14
# with regularization = 9+1+4 - 1.5*(1/log(3+1)+1/log(1+1)+1/log(2+1)) = 9.38
# alternative split: 'AAA|BAA' gives score = 9+3 - 1.5*(1/log(3+1)+1/log(3+1)) = 9.83
sequence = 'AAABAA'
splitter = pasio.SquareSplitter(
length_regularization_multiplier=1.5,
length_regularization_function=lambda x: 1 / np.log(1 + x))
optimal_split = splitter.split(sequence, SimpleScorer)
assert optimal_split[1] == [0, 3]
assert optimal_split[0] == 9 + 3 - 1.5 * (1 / np.log(3 + 1) +
1 / np.log(3 + 1))
# limiting possible splits
splitter = pasio.SquareSplitter(
length_regularization_multiplier=1.5,
length_regularization_function=lambda x: 1 / np.log(1 + x))
optimal_split = splitter.split(sequence,
SimpleScorer,
split_candidates=np.array([0, 4, 5]))
assert optimal_split[1] == [0, 4]
assert optimal_split[0] == 4 + 4 - 1.5 * (1 / np.log(4 + 1) +
1 / np.log(2 + 1))
def test_split_into_segments_candidates():
sequence = 'AAABBB'
optimal_split = pasio.SquareSplitter().split(
sequence, simple_scorer_factory, split_candidates=[0, 1, 2, 3, 5, 6])
assert optimal_split[1] == [0, 3]
assert optimal_split[0] == 9 + 9
sequence = 'AAABBB'
optimal_split = pasio.SquareSplitter().split(sequence,
simple_scorer_factory,
split_candidates=[0, 3, 5, 6])
assert optimal_split[1] == [0, 3]
assert optimal_split[0] == 9 + 9
sequence = 'AAABBBC'
optimal_split = pasio.SquareSplitter().split(sequence,
simple_scorer_factory,
split_candidates=[0, 3, 7])
assert optimal_split[1] == [0, 3]
assert optimal_split[0] == 9 + 4
sequence = 'AAABBBC'
optimal_split = pasio.SquareSplitter().split(sequence,
simple_scorer_factory,
split_candidates=[0, 3])
assert optimal_split[1] == [0, 3]
assert optimal_split[0] == 9 + 4
sequence = 'AAAAAA'
optimal_split = pasio.SquareSplitter().split(sequence,
simple_scorer_factory,
split_candidates=[0, 3])
assert optimal_split[1] == [0]
assert optimal_split[0] == 36
def test_split_into_segments_square():
sequence = 'A'
optimal_split = pasio.SquareSplitter().split(sequence,
simple_scorer_factory)
assert optimal_split[1] == [0]
assert optimal_split[0] == 1
sequence = 'AAA'
optimal_split = pasio.SquareSplitter().split(sequence,
simple_scorer_factory)
assert optimal_split[1] == [0]
assert optimal_split[0] == 9
sequence = 'AAABBB'
optimal_split = pasio.SquareSplitter().split(sequence,
simple_scorer_factory)
assert optimal_split[1] == [0, 3]
assert optimal_split[0] == 9 + 9
sequence = 'AAABBBC'
optimal_split = pasio.SquareSplitter().split(sequence,
simple_scorer_factory)
assert optimal_split[1] == [0, 3, 6]
assert optimal_split[0] == 9 + 9 + 1
sequence = 'ABBBC'
optimal_split = pasio.SquareSplitter().split(sequence,
simple_scorer_factory)
assert optimal_split[1] == [0, 1, 4]
assert optimal_split[0] == 1 + 9 + 1
def test_stat_split_into_segments_square():
def split_on_two_segments_or_not(counts, scorer_factory):
scorer = scorer_factory(counts)
best_score = scorer.score(0, len(counts))
split_point = 0
for i in range(len(counts)):
current_score = scorer.score(stop=i) + scorer.score(start=i)
if current_score > best_score:
split_point = i
best_score = current_score
return best_score, split_point
np.random.seed(4)
scorer_factory = lambda counts, split_candidates=None: pasio.LogMarginalLikelyhoodComputer(
counts, 1, 1, split_candidates)
for repeat in range(5):
counts = np.concatenate(
[np.random.poisson(15, 100),
np.random.poisson(20, 100)])
optimal_split = pasio.SquareSplitter().split(counts, scorer_factory)
two_split = split_on_two_segments_or_not(counts, scorer_factory)
assert optimal_split[0] >= two_split[0]
assert two_split[1] in optimal_split[1]
assert np.allclose(
optimal_split[0],
pasio.compute_score_from_splits(counts, optimal_split[1],
scorer_factory))
if (two_split[1] is None):
assert optimal_split[1] == [0, 200]
else:
assert abs(two_split[1] - 100) < 10
def test_split_into_segments_slidingwindow():
A = 'AAAAAAAAAAAAAAAA'
B = 'BBBBBBBBBBBBBBBBB'
sequence = A + B
splitter = pasio.SlidingWindowSplitter(
window_size=10, window_shift=5, base_splitter=pasio.SquareSplitter())
splits = splitter.split(sequence, simple_scorer_factory)
assert splits[1] == [0, len(A)]
assert splits[0] == len(A)**2 + len(B)**2
splitter = pasio.SlidingWindowSplitter(
window_size=10,
window_shift=5,
base_splitter=pasio.SquareSplitter(
split_number_regularization_multiplier=2))
splits = splitter.split(sequence, simple_scorer_factory)
assert splits[1] == [0, len(A)]
assert splits[0] == len(A)**2 + len(B)**2 - 2
def test_split_with_split_num_regularization():
# score of split 'AAA|B|AA' = 9+1+4 = 14
# with regularization = 9+1+4 - 3*2 = 8
# alternative split: 'AAA|BAA' gives score = 9+3-3*1 = 9
sequence = 'AAABAA'
splitter = pasio.SquareSplitter(
split_number_regularization_multiplier=3,
split_number_regularization_function=lambda x: x)
optimal_split = splitter.split(sequence, SimpleScorer)
assert optimal_split[1] == [0, 3]
assert optimal_split[0] == 9
def segmentation(counts, scorer, candidates=None):
optimal_split = pasio.SquareSplitter().split(counts,
scorer,
split_candidates=candidates)