def compute_dist(self, images, labels, sample_size=None):
"""Compute distances for pairs
sample_size: None or integer, number of pairs as all pairs can be large number.
If None, by default all possible pairs are considered.
Returns:
dist_embed: array of distances
actual_issame: array of booleans, True if positive pair, False if negative pair
"""
# Run forward pass to calculate embeddings
print('Generating pairs and computing embeddings...')
# Define generator to loop over data
gen = ImageDataGenerator(
data_format=K.image_data_format(),
preprocessing_function=self.backend_class.normalize,
)
features_shape = self.model.get_output_shape_at(0)[1:]
if sample_size is None:
n_pairs = comb(images.shape[0], 2, exact=True)
else:
n_pairs = int(sample_size)
embeddings = np.zeros(shape=(2, n_pairs) + features_shape)
actual_issame = np.full(shape=(n_pairs, ),
fill_value=True,
dtype=np.bool)
# Create all possible combinations of images (no repeats)
idx = 0
for i in range(images.shape[0]):
for j in range(i):
img_1 = gen.random_transform(images[i].astype(K.floatx()))
img_1 = gen.preprocessing_function(img_1)
img_2 = gen.random_transform(images[j].astype(K.floatx()))
img_2 = gen.preprocessing_function(img_2)
embeddings[0, idx] = self.model.predict_on_batch(
np.expand_dims(img_1, 0))
embeddings[1, idx] = self.model.predict_on_batch(
np.expand_dims(img_2, 0))
if labels[i] != labels[j]:
actual_issame[idx] = False
idx += 1
if idx >= n_pairs:
break
if idx >= n_pairs:
break
print('Number of pairs in evaluation {}, number of positive {}'.format(
len(actual_issame), np.sum(actual_issame)))
dist_emb = distance(embeddings[0], embeddings[1], distance_metric=0)
return dist_emb, actual_issame
def compute_dist(self, images, labels, sample_size=None):
"""Compute distances for pairs
sample_size: None or integer, number of pairs as all pairs can be large number.
If None, by default all possible pairs are considered.
"""
# Run forward pass to calculate embeddings
print('Generating pairs and computing embeddings...')
# Define generator to loop over data
gen = ImageDataGenerator(
data_format=K.image_data_format(),
preprocessing_function=self.backend_class.normalize,
)
if sample_size is None:
n_pairs = comb(images.shape[0], 2, exact=True)
else:
n_pairs = int(sample_size)
print('Computing distances for {} pairs...'.format(n_pairs))
distances = np.zeros(shape=(n_pairs, ))
actual_issame = np.full(shape=(n_pairs, ),
fill_value=True,
dtype=np.bool)
# Create all possible combinations of images (no repeats)
idx = 0
for i in range(images.shape[0]):
for j in range(i):
# Preprocess each image
img_1 = gen.random_transform(images[i].astype(K.floatx()))
img_1 = gen.preprocessing_function(img_1)
img_2 = gen.random_transform(images[j].astype(K.floatx()))
img_2 = gen.preprocessing_function(img_2)
# Predict distance using Siamese model
distances[idx] = self.model.predict_on_batch(
[np.expand_dims(img_1, 0),
np.expand_dims(img_2, 0)])
# Change flag to False for negative pairs
if labels[i] != labels[j]:
actual_issame[idx] = False
if idx > 0 and idx % 100 == 0:
print('Computed distances for {} pairs'.format(idx))
# print('Distance: {:.2f} actual: {}. labels {}, {}'.format(distances[idx], actual_issame[idx],
# labels[i], labels[j]))
idx += 1
if idx >= n_pairs:
break
if idx >= n_pairs:
break
print('Number of pairs in evaluation {}, number of positive {}'.format(
len(actual_issame), np.sum(actual_issame)))
return distances, actual_issame
