def test_depth_dist():
"""
test for depth_dist function
"""
X_train = np.concatenate(
(np.random.normal(loc=(1, 2), scale=.6, size=(100, 2)),
np.random.normal(loc=(-1.2, -.5), scale=.6, size=(100, 2))),
axis=0)
y_train = np.concatenate((np.zeros(100,
dtype=np.int), np.ones(100,
dtype=np.int)))
amount = np.int(200)
# creating a random forest
rf_class_known = sklearn.ensemble.RandomForestClassifier(
n_estimators=1, min_samples_leaf=1)
random_forest = rf_class_known.fit(X=np.array(X_train)[:amount, :],
y=y_train[:amount].ravel())
data = np.array(X_train[:amount, :])
depth_dict, max_depth = smooth_rf.calc_depth_for_forest(random_forest,
verbose=False)
Vt_dict = smooth_rf.make_Vt_mat(random_forest,
data,
depth_dict=depth_dict,
verbose=False)
Ut_prime_dict = smooth_rf.make_Ut_prime_mat_no_sym(Vt_dict,
Vt_dict,
max_depth=max_depth,
verbose=False)
Ut_prime_dict = smooth_rf.remove_0_from_Ut_prime(Ut_prime_dict)
if len(Ut_prime_dict) > 0:
K_mat = smooth_rf.make_kernel(Ut_prime_dict)
DD_mat = smooth_rf.depth_dist(K_mat)
assert K_mat.shape == DD_mat.shape, \
"dimensions between K_mat and DD_mat should be the same"
if type(DD_mat) is sparse.coo.core.COO:
assert np.all(np.diag(DD_mat.todense()) == 0), \
"diagonal should be naturally 0 (has error)"
else:
assert np.all(np.diag(DD_mat) == 0), \
"diagonal should be naturally 0 (has error)"
assert np.all(DD_mat >= 0), \
"all entries should be positive in DD (has error)"
else:
K_mat = smooth_rf.make_kernel(Ut_prime_dict)
assert K_mat == 0, \
"when you provide an empty Ut_prime_dict you should get a 0"
def test_make_kernel():
"""
test for make_kernel
"""
X_train = np.concatenate(
(np.random.normal(loc=(1, 2), scale=.6, size=(100, 2)),
np.random.normal(loc=(-1.2, -.5), scale=.6, size=(100, 2))),
axis=0)
y_train = np.concatenate((np.zeros(100,
dtype=np.int), np.ones(100,
dtype=np.int)))
amount = np.int(200)
# creating a random forest
rf_class_known = sklearn.ensemble.RandomForestClassifier(
n_estimators=1, min_samples_leaf=1)
random_forest = rf_class_known.fit(X=np.array(X_train)[:amount, :],
y=y_train[:amount].ravel())
data = np.array(X_train[:amount, :])
depth_dict, max_depth = smooth_rf.calc_depth_for_forest(random_forest,
verbose=False)
Vt_dict = smooth_rf.make_Vt_mat(random_forest,
data,
depth_dict=depth_dict,
verbose=False)
Ut_prime_dict = smooth_rf.make_Ut_prime_mat_no_sym(Vt_dict,
Vt_dict,
max_depth=max_depth,
verbose=False)
Ut_prime_dict = smooth_rf.remove_0_from_Ut_prime(Ut_prime_dict)
if len(Ut_prime_dict) > 0:
K_mat = smooth_rf.make_kernel(Ut_prime_dict)
assert K_mat.shape[0:2] == (K_mat.shape[0],K_mat.shape[0]), \
"returned K matrix is not symmetric when the inputs were."
assert len(K_mat.shape) == 2, \
"returned K matrix is not 2d as expected."
else:
K_mat = smooth_rf.make_kernel(Ut_prime_dict)
assert K_mat == 0, \
"when you provide an empty Ut_prime_dict you should get a 0"
def test_make_Vt_mat():
"""
tests for make_Vt_mat
"""
X_train = np.concatenate(
(np.random.normal(loc=(1, 2), scale=.6, size=(100, 2)),
np.random.normal(loc=(-1.2, -.5), scale=.6, size=(100, 2))),
axis=0)
y_train = np.concatenate((np.zeros(100,
dtype=np.int), np.ones(100,
dtype=np.int)))
amount = np.int(200)
s = 20
c = y_train[:amount]
# creating a random forest
rf_class_known = sklearn.ensemble.RandomForestClassifier(
n_estimators=100, min_samples_leaf=1)
fit_rf_known = rf_class_known.fit(X=np.array(X_train)[:amount, :],
y=y_train[:amount].ravel())
random_forest = fit_rf_known
data = np.array(X_train[:amount, :])
Vt_dict = smooth_rf.make_Vt_mat(random_forest, data, verbose=False)
assert len(Vt_dict) == 100, \
"incorrect number of trees suggested in the full Vt output"
for _ in range(10):
r_idx = np.random.randint(100)
random_Vt_dict = Vt_dict[r_idx]
assert type(random_Vt_dict) == dict, \
"output of _make_Vt_mat_tree is not a dictionary"
assert np.all([x.shape[0] == data.shape[0] \
for x in random_Vt_dict.values()]), \
"output of _make_Vt_mat_tree elements have incorrect number of rows"
assert np.sum([x.shape[1] for x in random_Vt_dict.values()]) == \
len(random_forest.estimators_[r_idx].tree_.children_left), \
"output of split of Vt matrices have more columns than a" + \
" full Vt mat would"