def filter_candidate_pool(self, candidate_pool, char):
"""Filter candidate pool based on secondary distance metrics and
selected threshold. Secondary distance is calculated between the image
of the anchor character and the image of the candidate character.
Args:
candidate_pool: Set of Char, the set of possible confusables.
char: Char, single character whose corresponding image must exists
in self.img_dir.
"""
# Get secondary distance metrics
image_metrics = Distance(self.img_format).get_metrics()
if self.secondary_distance_type not in image_metrics.keys():
raise ValueError(
"Expect secondary_distance_type to be one of {}.".format(
image_metrics.keys()))
secondary_dis = image_metrics[self.secondary_distance_type]
# Filter candidate pool to get confusables
confusables = []
for candidate in candidate_pool:
if ord(char) == ord(candidate):
continue
dis = calculate_from_path(secondary_dis, self._label_img_map[char],
self._label_img_map[candidate])
if dis <= self.secondary_filter_threshold:
confusables.append((candidate, dis))
return confusables
def test_emb_metrics(self):
# Test embedding metrics
dis = Distance(ImgFormat.EMBEDDINGS)
metrics = dis.get_metrics()
# Test manhattan distance
self.assertEqual(metrics.manhattan(self.emb_0, self.emb_1), 3)
self.assertEqual(metrics.manhattan(self.emb_1, self.emb_1), 0)
self.assertAlmostEqual(metrics.manhattan(self.emb_123, self.emb_1), 3)
self.assertAlmostEqual(metrics.manhattan(self.emb_0, self.emb_123), 6)
# Test euclidean distance
self.assertAlmostEqual(metrics.euclidean(self.emb_0, self.emb_1),
np.sqrt(3))
self.assertEqual(metrics.euclidean(self.emb_1, self.emb_1), 0)
self.assertAlmostEqual(metrics.euclidean(self.emb_123, self.emb_0),
np.sqrt((1**2) + (2**2) + (3**2)))
# Test exception
with self.assertRaises(TypeError):
metrics.manhattan(self.emb_0.tolist(), self.emb_1)
with self.assertRaises(ValueError):
metrics.manhattan(self.emb_0, np.ones(4))
with self.assertRaises(TypeError):
metrics.euclidean(self.emb_0.tolist(), self.emb_1)
with self.assertRaises(ValueError):
metrics.euclidean(self.emb_0, np.ones(4))
def get_candidate_pool_for_char(self, char):
"""Obtain the candidates for confusables for specified 'char'. Use
the reduced representations generated by PCA and select candidates based
on primary distance metrics.
Args:
char: Char, single character, must exists in self.labels.
Returns:
candidate_pool: Set of Char, the set of possible confusables.
candidate_dis: Dict, mapping from candidates to their respective
distances.
"""
# Get character index in labels and embeddings
idx = self.labels.index(char)
# Get a pool of possible candidates for secondary filter
candidate_pool = set()
# Store distances between all confusables and anchor
candidate_dis = dict()
for embs in self._reps:
# Get embedding anchor to compare with others
emb_anchor = embs[idx]
# Get primary distance metrics
embedding_metrics = Distance(ImgFormat.EMBEDDINGS).get_metrics()
if self.primary_distance_type not in embedding_metrics.keys():
raise ValueError(
"Expect primary_distance_type to be one of {}.".format(
embedding_metrics.keys()))
primary_dis = embedding_metrics[self.primary_distance_type]
# Get distance from anchor embedding to all other embeddings
distances = []
for emb in embs:
distances.append(primary_dis(emb_anchor, emb))
label_dis_pairs = list(zip(self.labels, distances))
# Get top n candidates using the primary distance metric
top_n = []
for label, dis in label_dis_pairs:
if len(top_n) < self.n_candidates:
# Append reversed tuple for sorting
bisect.insort(top_n, (dis, label))
else:
if dis < top_n[self.n_candidates - 1][0]:
# If the distance is lower than the largest of the
# candidates we only keep top N
bisect.insort(top_n, (dis, label))
top_n = top_n[:self.n_candidates - 1]
# Store all candidate distances
candidate_dis["PCA" + str(embs.shape[1])] = top_n
candidate_pool = candidate_pool.union(
set([entry[1] for entry in top_n]))
return candidate_pool, candidate_dis
def test_rgb_metrics(self):
# Test RGB metrics
dis = Distance(ImgFormat.RGB)
metrics = dis.get_metrics()
# Test manhattan distance
self.assertEqual(metrics.manhattan(self.img_rgb_0, self.img_rgb_255),
255.0)
self.assertEqual(metrics.manhattan(self.img_rgb_255, self.img_rgb_0),
255.0)
self.assertAlmostEqual(
metrics.manhattan(self.img_rgb_topleft, self.img_rgb_botright),
(abs(255 - 200) + abs(255 - 100)) / (3 * 3))
self.assertEqual(metrics.manhattan(self.img_rgb_255, self.img_rgb_255),
0)
# Test sum squared distance
self.assertEqual(metrics.sum_squared(self.img_rgb_0, self.img_rgb_255),
1.0)
self.assertEqual(
metrics.sum_squared(self.img_rgb_0, self.img_rgb_topleft), 1.0)
self.assertAlmostEqual(
metrics.sum_squared(self.img_rgb_topleft, self.img_rgb_botright),
0.049633607)
self.assertAlmostEqual(
metrics.sum_squared(self.img_rgb_255, self.img_rgb_botright),
0.005283143)
self.assertEqual(
metrics.sum_squared(self.img_rgb_255, self.img_rgb_255), 0)
# Test cross correlation distance
self.assertEqual(
metrics.cross_correlation(self.img_rgb_0, self.img_rgb_255), 0)
self.assertAlmostEqual(
metrics.cross_correlation(self.img_rgb_topleft,
self.img_rgb_botright), 0.9755619)
self.assertAlmostEqual(
metrics.cross_correlation(self.img_rgb_255, self.img_rgb_botright),
0.99759716)
self.assertEqual(
metrics.cross_correlation(self.img_rgb_255, self.img_rgb_255), 1.0)
# Test exception
with self.assertRaises(TypeError):
metrics.manhattan(self.img_rgb_0.tolist(), self.img_rgb_255)
with self.assertRaises(ValueError):
metrics.manhattan(self.img_gray_0, self.img_rgb_255)
with self.assertRaises(TypeError):
metrics.sum_squared(self.img_rgb_0.tolist(), self.img_rgb_255)
with self.assertRaises(ValueError):
metrics.sum_squared(self.img_gray_0, self.img_rgb_255)
with self.assertRaises(TypeError):
metrics.cross_correlation(self.img_rgb_0.tolist(),
self.img_rgb_255)
with self.assertRaises(ValueError):
metrics.cross_correlation(self.img_gray_0, self.img_rgb_255)
def test_img_format_setter(self):
# Test setter in initialization
dis = Distance(img_format=ImgFormat.A8)
self.assertEqual(dis._img_format, ImgFormat.A8)
# Test setter after initialization
dis.img_format = ImgFormat.EMBEDDINGS
self.assertEqual(dis._img_format, ImgFormat.EMBEDDINGS)
# Test exception
with self.assertRaises(TypeError):
dis.img_format = 5
def test_default_init(self):
"""Test default initialization. When default initialization value
changes, or any private attribute does not match public attribute, this
test will fail."""
dis = Distance()
self.assertEqual(dis._img_format, ImgFormat.RGB)