def test_init_(self) -> None:
"""
Test if implementation is usable.
"""
impl = OcclusionScoring()
assert impl.is_usable() and isinstance(
impl, ImageClassifierSaliencyMapGenerator)
def test_standard_1class_scores(self) -> None:
"""
Test basic scoring on known values and non-square masks.
"""
impl = OcclusionScoring()
# Three Pertrubation masks of size 4 x 6 for 1 class
image_confs_1_class_ = np.array([0.6])
pertb_confs_1_class_ = np.array([[0.3], [0.65], [0.12], [0.18], [0.36],
[0.42]])
sal = impl.generate(image_confs_1_class_, pertb_confs_1_class_,
EXPECTED_MASKS_4x6)
standard_sal = np.load(os.path.join(DATA_DIR, 'OccScorSal.npy'))
assert sal.shape == (1, 4, 6)
assert np.allclose(standard_sal, sal)
def test_bad_alignment_confs(self) -> None:
"""
Test that a mismatch of input ref-image and perturbed-image confidence
classes causes the expected exception.
"""
test_ref_confs = np.ones([3]) # ONE MORE than pert conf mat.
test_pert_confs = np.ones((4, 2))
test_pert_masks = np.ones((4, 3, 3))
impl = OcclusionScoring()
with pytest.raises(
ValueError,
match=r"Number of classes in original image and perturbed image "
r"do not match"):
impl.generate(test_ref_confs, test_pert_confs, test_pert_masks)
def test_20class_scores(self) -> None:
"""
Test scoring for 20 classes.
"""
impl = OcclusionScoring()
np.random.seed(2)
# Three Pertrubation masks of height and width 10px for 20 classes
image_confs_1_class_ = np.random.rand(20)
pertb_confs_1_class_ = np.random.rand(3, 20)
mask_confs_1_class_ = np.random.randint(low=0,
high=2,
size=(3, 10, 10),
dtype='int')
sal = impl.generate(image_confs_1_class_, pertb_confs_1_class_,
mask_confs_1_class_)
assert sal.shape == (20, 10, 10)
def test_1class_scores(self) -> None:
"""
Test basic scoring with a single class for broadcasting sanity check.
"""
impl = OcclusionScoring()
np.random.seed(2)
# Three Pertrubation masks of height and width 10px for 1 class
image_confs_1_class_ = np.random.rand(1)
pertb_confs_1_class_ = np.random.rand(3, 1)
mask_confs_1_class_ = np.random.randint(low=0,
high=2,
size=(3, 10, 10),
dtype='int')
sal = impl.generate(image_confs_1_class_, pertb_confs_1_class_,
mask_confs_1_class_)
assert sal.shape == (1, 10, 10)
def test_bad_alignment_masks(self) -> None:
"""
Test that the number of input perturbed image confidences and masks
match.
"""
test_ref_confs = np.ones([2])
test_pert_confs = np.ones((4, 2))
# Different number of masks from confs
test_pert_masks = np.ones((7, 3, 3))
impl = OcclusionScoring()
with pytest.raises(
ValueError,
match=r"Number of perturbation masks and respective "
r"confidence lengths do not match"):
impl.generate(test_ref_confs, test_pert_confs, test_pert_masks)