def test__compute_single_preset_relevance_bounds(randomstate):
problem = ProblemName.CLASSIFICATION
data = quick_generate(
problem,
n_samples=300,
n_features=4,
n_redundant=2,
n_strel=2,
random_state=randomstate,
)
X_orig, y = data
X = scale(X_orig)
model = FRI(problem, random_state=randomstate, n_jobs=1)
model.fit(X, y)
normal_range = model.interval_.copy()
i = 0
preset = [model.interval_[i, 0]]
range = model._relevance_bounds_computer.compute_single_preset_relevance_bounds(
i, preset)
assert normal_range.shape == range.shape
assert range[i][0] == preset[0]
assert range[i][1] == preset
preset = model.interval_[i, 0]
range = model._relevance_bounds_computer.compute_single_preset_relevance_bounds(
i, preset)
assert normal_range.shape == range.shape
assert range[i][0] == preset
assert range[i][1] == preset
def test_lupi_model(problem, random_state):
model = FRI(problem, random_state=random_state)
X, X_p, y = quick_generate(problem, random_state=random_state)
combined = np.hstack([X, X_p])
model.fit(combined, y, lupi_features=X.shape[1])
assert len(model.allrel_prediction_) == combined.shape[1]
def test_normal_model(problem, random_state):
model = FRI(problem, random_state=random_state)
X, y = quick_generate(problem, random_state=random_state)
model.fit(X, y)
assert len(model.allrel_prediction_) == X.shape[1]
def test_normal_model(random_state):
problem = fri.ProblemName.CLASSIFICATION
model = FRI(problem, random_state=random_state)
X, y = quick_generate(problem, random_state=random_state, n_features=5)
model.fit(X, y)
assert len(model.allrel_prediction_) == X.shape[1]
groups, links = model.get_grouping()
print(groups)
assert len(groups) == X.shape[1]
def test__compute_multi_preset_relevance_bounds(problem, randomstate):
data = quick_generate(
problem,
n_samples=300,
n_features=4,
n_redundant=2,
n_strel=2,
random_state=randomstate,
)
X_orig, y = data
X = scale(X_orig)
model = FRI(problem, random_state=randomstate, n_jobs=1)
model.fit(X, y)
normal_range = model.interval_.copy()
# We fix two feature values at the same time.
# Note that we can not use the relevance bounds for multiple features, because of infeasibility
# In the singular optimization case we exhaust the complete slack of our model constraints
# If whe use multiple relevance bounds we double count this slack and the model optimization rightfully claims infeasible
# Instead we use the mean value of the relevance bounds, which should be working in most cases
i = 0
mean_0 = np.mean([model.interval_[i, 0], model.interval_[i, 1]])
preset_0 = [mean_0, mean_0]
i = 1
mean_1 = np.mean([model.interval_[i, 0], model.interval_[i, 1]])
preset_1 = [mean_1, mean_1]
presetModel = {0: preset_0, 1: preset_1}
range = model._relevance_bounds_computer.compute_multi_preset_relevance_bounds(
presetModel)
assert normal_range.shape == range.shape
i = 0
assert range[i][0] == preset_0[0]
assert range[i][1] == preset_0[1]
i = 1
assert range[i][0] == preset_1[0]
assert range[i][1] == preset_1[1]
def test_quick_generate(problem):
data = quick_generate(problem)
assert (
len(data) == 2 or len(data) == 3
) # Check if two-tuple or three-tuple depending on if we have lupi data or not
def data(prob):
return quick_generate(prob)