def test_est_tuning():
"""Check that est_tuning returns the estimator with minimum auc."""
for algo in algorithms:
name_algo = algo.name
parameters = algo_param[name_algo]
param_grid = ParameterGrid(parameters)
alphas = rng.randint(1, 100, size=5) / 100
alphas = np.sort(alphas)
clf_est, offsets_est = est_tuning(X_train, X_test, algo, param_grid,
alphas, U, vol_tot_cube)
# check that clf_est gives the minimum auc
score_function = clf_est.score_samples
vol_est, _ = _compute_volumes(score_function, alphas, X_test, U,
vol_tot_cube)
auc_est = auc(alphas, vol_est)
auc_algo = np.zeros(len(param_grid))
for p, param in enumerate(param_grid):
clf = algo(**param)
clf = clf.fit(X_train)
score_function_p = clf.score_samples
vol_p, _ = _compute_volumes(score_function_p, alphas, X_test, U,
vol_tot_cube)
auc_algo[p] = auc(alphas, vol_p)
assert_equal(np.min(auc_algo), auc_est)
clf_test = clf_est.score_samples(X_test)
proba_offsets_est = (clf_test >= offsets_est[:, np.newaxis])
# this test requires to have a large number of samples because
# np.percentile is an empirical quantile which uses interpolation.
# this is also why we ask the values to be equal only up to the
# second decimal.
assert_array_almost_equal(np.mean(proba_offsets_est, axis=1),
alphas,
decimal=2)
def test_compute_volumes_toy():
"""Check _compute_volumes on a toy scoring function."""
def score_function_toy(X):
""" score(x_0, x_1) = x_0 """
return X[:, 0]
# regular grid with step size equal to 0.01
xx, yy = np.meshgrid(np.arange(0, 1, 0.01), np.arange(0, 1, 0.01))
grid = np.c_[xx.ravel(), yy.ravel()]
alphas = rng.randint(1, 100, size=5) / 100
vols, offsets = _compute_volumes(score_function_toy, alphas, grid, grid,
1.)
# check values of volume
assert_array_equal(alphas, vols)
# check values of offsets
pred_offsets = (score_function_toy(grid) >= offsets[:, np.newaxis])
assert_array_equal(np.mean(pred_offsets, axis=1), alphas)
def test_compute_volumes():
"""Check _compute_volumes for several masses."""
estimators = [
AverageKLPE(k=3, novelty=True),
MaxKLPE(k=3, novelty=True),
OCSVM(sigma=1.),
IsolationForest(n_estimators=5, random_state=2),
KernelSmoothing()
]
alphas = rng.randint(1, 100, size=5) / 100
alphas = np.sort(alphas)
for clf in estimators:
clf = clf.fit(X_train)
clf_test = clf.score_samples(X_test)
min_test = np.min(clf_test)
max_test = np.max(clf_test)
score_function = clf.score_samples
vols, offsets = _compute_volumes(score_function, alphas, X_test, U,
vol_tot_cube)
# check increasing order of volumes and decreasing order of offsets
assert_array_equal(vols, np.sort(vols))
assert_array_equal(offsets, -np.sort(-offsets))
# check volumes in [0, vol_tot_cube]
assert_true(np.all(0 <= vols) and np.all(vols <= vol_tot_cube))
# check offset values
assert_true(
np.all(min_test <= offsets) and np.all(offsets <= max_test))
proba_offsets_pos = (clf_test >= offsets[:, np.newaxis])
# this test requires to have a large number of samples because
# np.percentile is an empirical quantile which uses interpolation.
# this is also why we ask the values to be equal only up to the
# second decimal.
assert_array_almost_equal(np.mean(proba_offsets_pos, axis=1),
alphas,
decimal=2)