def test_measurer_adjacent_2():
rule1 = Rule(
[
Statement(0, Relation.MT, 1),
Statement(0, Relation.LEQ, 3),
Statement(2, Relation.MT, 1),
Statement(2, Relation.LEQ, 3),
],
1,
)
rule2 = Rule(
[
Statement(0, Relation.MT, 1),
Statement(0, Relation.LEQ, 3),
Statement(1, Relation.MT, 1),
Statement(1, Relation.LEQ, 3),
],
1,
)
assert OverlappingMeasurer().measure(rule1,
rule2) == AdjacentOrNot.ADJACENT
assert OverlappingMeasurer().measure(rule2,
rule1) == AdjacentOrNot.ADJACENT
def bound_rule(rule: Rule, x_train) -> Rule:
feature_max_values = np.max(x_train, axis=0)
feature_min_values = np.min(x_train, axis=0)
new_statements_for_rule = set(rule.statements)
for statement in rule.statements:
new_statements_for_rule = new_statements_for_rule.union(
bound_if_needed(rule, statement, feature_min_values, feature_max_values)
)
for feature_idx in range(len(feature_max_values)):
statements_for_feature = rule.get_statements_for_feature(feature_idx)
if len(statements_for_feature) == 0:
lower_statement = Statement(
feature_idx, Relation.MT, feature_min_values[feature_idx] - EPS
)
upper_statement = Statement(
feature_idx, Relation.LEQ, feature_max_values[feature_idx]
)
new_statements_for_rule = new_statements_for_rule.union(
{lower_statement, upper_statement}
)
return Rule(new_statements_for_rule, rule.distribution_or_class)
def join_consecutive_statements(rule: Rule) -> Rule:
all_statements = set()
for feature in rule.get_features():
statements = rule.get_statements_for_feature(feature)
new_this_rule_statements = set(statements)
sorted_by_threshold = iter(sorted(statements, key=lambda s: s.threshold))
try:
current_statement = next(sorted_by_threshold)
while True:
next_statement = next(sorted_by_threshold)
if current_statement.relation == next_statement.relation:
if current_statement.relation == Relation.MT:
new_this_rule_statements.remove(next_statement)
else:
new_this_rule_statements.remove(current_statement)
current_statement = next_statement
except StopIteration:
pass
all_statements = all_statements.union(new_this_rule_statements)
return Rule(all_statements, rule.distribution_or_class)
def join_if_possible(rule1: Rule, rule2: Rule) -> Optional[Rule]:
spans_by_feature_idx_1 = {
feature: calculate_span_for_statements(statements)
for feature, statements in groupby(
rule1.statements, lambda statement: statement.feature_idx)
}
spans_by_feature_idx_2 = {
feature: calculate_span_for_statements(statements)
for feature, statements in groupby(
rule2.statements, lambda statement: statement.feature_idx)
}
overlaps = any([
any_span_overlaps(spans1, spans2)
for feature1, spans1 in spans_by_feature_idx_1.items()
for feature2, spans2 in spans_by_feature_idx_2.items()
if feature1 == feature2
])
if overlaps:
return None
new_distribution = defaultdict(lambda: 0)
for label, value in rule1.distribution_or_class.items():
new_distribution[label] = new_distribution[label] + value
for label, value in rule2.distribution_or_class.items():
new_distribution[label] = new_distribution[label] + value
return Rule(rule1.statements + rule2.statements, new_distribution)
def bound_if_needed(
rule: Rule, statement: Statement, feature_min_values, feature_max_values
) -> Set[Statement]:
statements = set(rule.get_statements_for_feature(statement.feature_idx)).difference(
{statement}
)
new_statements = set()
next_higher = pipe(
statements, filter(lambda s: s.threshold > statement.threshold), list
)
next_lower = pipe(
statements, filter(lambda s: s.threshold < statement.threshold), list
)
if statement.relation == Relation.LEQ and not any(next_lower):
new_statements.add(
Statement(
statement.feature_idx,
Relation.MT,
feature_min_values[statement.feature_idx],
)
)
elif statement.relation == Relation.MT and not any(next_higher):
new_statements.add(
Statement(
statement.feature_idx,
Relation.LEQ,
feature_max_values[statement.feature_idx] - EPS,
)
)
else:
return set()
return new_statements
def any_span_overlaps(spans1: Set[Span], spans2: Set[Span]) -> bool:
return any(
[span1.overlaps(span2) for span1, span2 in product(spans1, spans2)])
if (adjacency == AdjacentOrNot.ADJACENT
and rule1.classified_class != rule2.classified_class):
return None
return Rule(rule1.statements + rule2.statements, rule1.classified_class)
def to_rule(hyperrect: Hyperrectangle) -> Rule:
rule_statements = set()
for feature, statement in hyperrect.get_statement_by_feature().items():
if statement.upper_boundary != np.inf and statement.lower_boundary != -np.inf:
rule_statements.add(
Statement(feature.idx, Relation.LEQ, statement.upper_boundary))
rule_statements.add(
Statement(feature.idx, Relation.MT,
statement.lower_boundary - EPS))
return Rule(list(rule_statements), hyperrect.label)
def test_measurer_not_adjacent_real():
rule1 = Rule(
[
Statement(feature_idx=3, relation=Relation.LEQ, threshold=2.5),
Statement(feature_idx=2, relation=Relation.MT, threshold=1.1),
Statement(feature_idx=0, relation=Relation.MT, threshold=5.75),
Statement(feature_idx=1,
relation=Relation.LEQ,
threshold=3.700000047683716),
Statement(feature_idx=0, relation=Relation.LEQ, threshold=7.9),
Statement(feature_idx=2,
relation=Relation.LEQ,
threshold=4.950000047683716),
Statement(feature_idx=3,
relation=Relation.MT,
threshold=1.699999988079071),
Statement(feature_idx=1, relation=Relation.MT, threshold=2.0),
],
0,
)
rule2 = Rule(
[
Statement(feature_idx=3, relation=Relation.LEQ, threshold=2.5),
Statement(feature_idx=2, relation=Relation.LEQ, threshold=6.9),
Statement(feature_idx=2,
relation=Relation.MT,
threshold=2.350000023841858),
Statement(feature_idx=3,
relation=Relation.MT,
threshold=1.6500000357627869),
],
1,
)
assert OverlappingMeasurer().measure(rule1,
rule2) == AdjacentOrNot.ADJACENT
assert OverlappingMeasurer().measure(rule2,
rule1) == AdjacentOrNot.ADJACENT
def test_measures_1():
rule = Rule(
statements=[Statement(0, Relation.LEQ, threshold=5)], distribution_or_class=0
)
x = [[0], [1], [2], [6], [7], [8]]
y = [0, 0, 1, 0, 1, 0]
measures = BayesianRuleMeasures.create(rule, x, y)
assert measures.a() == 2
assert measures.b() == 2
assert measures.c() == 1
assert measures.d() == 1
def test_get_rule_corners_not_sorted_statements():
r1 = Rule(
[
Statement(1, Relation.LEQ, 4),
Statement(1, Relation.MT, 8),
Statement(0, Relation.LEQ, 2),
Statement(0, Relation.MT, 0),
],
0,
)
edges = get_rule_edges(r1)
assert {(2, 4), (0, 4), (2, 8), (0, 8)} == edges
def as_rules(self, rects: Collection[Rectangle]) -> Collection[Rule]:
rules = []
for rect in rects:
statements = []
for feature, bound in rect.feature_by_bounds.items():
statements.append(Statement(feature, Relation.MT, bound.lower))
statements.append(Statement(feature, Relation.LEQ,
bound.upper))
labeling = (self.rect_by_labeling[rect] if rect
in self.rect_by_labeling else self.default_class)
rules.append(Rule(statements, labeling))
return rules
def to_nnge_hyper(rule: Rule, x) -> Hyperrectangle:
hyperrect_statemets = set()
for feature_idx, statements in rule.get_statements_by_feature().items():
if len(statements) != 2:
raise Exception("nah")
sorted_thresholds = pipe(statements, map(lambda s: s.threshold),
sorted, list)
feature = HyperrectFeature(feature_idx, FeatureType.REAL)
hyperrect_statemets.add(
HyperrectStatement(
feature,
lower_boundary=sorted_thresholds[0] - EPS,
upper_boundary=sorted_thresholds[1],
))
hyper = Hyperrectangle(hyperrect_statemets, label=rule.classified_class)
return hyper
def test_bound_rule():
rule = Rule(
[
Statement(0, Relation.MT, 0),
Statement(1, Relation.LEQ, 20),
],
1,
)
x_train = [[-5, -5], [30, 30]]
bounded_rule = bound_rule(rule, x_train)
assert set(bounded_rule.statements) == {
Statement(0, Relation.MT, 0),
Statement(0, Relation.LEQ, 30),
Statement(1, Relation.LEQ, 20),
Statement(1, Relation.MT, -5 - EPS),
}
def test_join_consecutive():
rule = Rule(
[
Statement(0, Relation.MT, 10),
Statement(0, Relation.MT, 20),
Statement(0, Relation.LEQ, 35),
Statement(0, Relation.LEQ, 40),
Statement(0, Relation.LEQ, 50),
Statement(1, Relation.MT, 10),
Statement(1, Relation.MT, 20),
],
1,
)
actual = join_consecutive_statements(rule)
assert set(actual.statements) == {
Statement(0, Relation.MT, 10),
Statement(0, Relation.LEQ, 50),
Statement(1, Relation.MT, 10),
}
def recurse(node: int, statements: List[Statement] = []) -> Set[Rule]:
currentNodeFeatureIndex: int = tree_.feature[node]
if currentNodeFeatureIndex != _tree.TREE_UNDEFINED:
threshold: float = tree_.threshold[node]
leftStatements = statements + [
Statement(currentNodeFeatureIndex, Relation.LEQ, threshold)
]
leftRules = recurse(tree_.children_left[node], leftStatements)
rightStatements = statements + [
Statement(currentNodeFeatureIndex, Relation.MT, threshold)
]
rightRules = recurse(tree_.children_right[node], rightStatements)
return leftRules.union(rightRules)
else:
samplesCountForEachClass = tree_.value[node][0]
samplesCountByClass = {
self._skLearnTree.classes_[idx]: count
for idx, count in enumerate(samplesCountForEachClass)
}
return {Rule(set(statements), samplesCountByClass)}
return any(
[span1.overlaps(span2) for span1, span2 in product(spans1, spans2)])
if (adjacency == AdjacentOrNot.ADJACENT
and rule1.classified_class != rule2.classified_class):
return None
return Rule(rule1.statements + rule2.statements, rule1.classified_class)
@pytest.mark.skip("to fix")
@pytest.mark.parametrize(
"rule1,rule2,expected",
[
(
Rule([Statement(0, Relation.MT, 0),
Statement(0, Relation.LEQ, 5)], 1),
Rule([Statement(1, Relation.MT, 0),
Statement(1, Relation.LEQ, 5)], 1),
Rule(
[
Statement(0, Relation.MT, 0),
Statement(0, Relation.LEQ, 5),
Statement(1, Relation.MT, 0),
Statement(1, Relation.LEQ, 5),
],
1,
),
),
(
Rule([Statement(0, Relation.MT, 0)], 1),
Rule([Statement(1, Relation.MT, 0)], 2),