def embedding_centers_to_logits(embedding,
centers,
similarity_strategy,
max_value=10.0):
"""logits from embedding and centers based on distance to centers.
Args:
embedding: A tf.float32 tensor of size [N, embedding_size].
centers: A tf.float32 tensor of size [k, embedding_size].
similarity_strategy: Strategy for computing the similarity between
embedding vectors.
max_value: Maximum absolute value of any dimension of the logits.
Returns:
A tf.float32 tensor of [k, N] that contains k logit maps.
Raises:
ValueError: If similarity_strategy is unknown.
"""
if similarity_strategy == 'dotproduct':
logits = tf.matmul(centers, embedding, transpose_b=True)
elif similarity_strategy == 'distance':
logits = -instance_segmentation_utils.inputs_distances_to_centers(
centers, embedding)
else:
raise ValueError('Similarity strategy is unknown')
return tf.minimum(tf.maximum(logits, -max_value), max_value)
def body(i, init_samples, indices, min_distances):
"""while loop body. Pick the next center given previously picked centers.
Args:
i: For loop index.
init_samples: Initial samples that is modified inside the body.
indices: Indices of picked samples.
min_distances: Minimum distance of the embedding to centers so far.
Returns:
i and init_centers.
"""
best_new_sample_ind = tf.cast(
tf.math.argmax(
scores_coef * log_scores +
tf.math.log(min_distances + sys.float_info.min),
axis=0),
dtype=tf.int32)
indices += tf.pad(tf.stack([best_new_sample_ind]),
paddings=[[i, num_samples-i-1]])
init_samples_i = tf.expand_dims(tf.gather(inputs, best_new_sample_ind), 0)
init_samples += tf.pad(init_samples_i,
paddings=tf.stack([tf.stack([i, num_samples-i-1]),
tf.stack([0, 0])]))
min_distances = tf.minimum(
tf.squeeze(
instance_segmentation_utils.inputs_distances_to_centers(
inputs, init_samples_i),
axis=1), min_distances)
i += 1
return i, init_samples, indices, min_distances
def test_inputs_Distances_to_centers(self):
inputs = tf.random.uniform(
[100, 8], minval=-10, maxval=10.0, dtype=tf.float32)
centers = tf.random.uniform(
[5, 8], minval=-10, maxval=10.0, dtype=tf.float32)
distances1 = isu.inputs_distances_to_centers(inputs, centers)
num_centers = tf.shape(centers)[0]
inputs_reshaped = tf.tile(tf.expand_dims(inputs, axis=1),
tf.stack([1, num_centers, 1]))
distances2 = tf.reduce_sum(tf.square(inputs_reshaped - centers), axis=2)
self.assertAllClose(distances1.numpy(), distances2.numpy(), atol=0.001)
def embedding_centers_to_soft_masks_by_distance(embedding, centers):
"""Masks from embedding and centers based on distance to centers.
Args:
embedding: A tf.float32 tensor of size [N, embedding_size].
centers: A tf.float32 tensor of size [k, embedding_size].
Returns:
A tf.float32 tensor of [k, N] that contains k soft masks.
"""
dists_to_centers = instance_segmentation_utils.inputs_distances_to_centers(
centers, embedding)
return tf.nn.sigmoid(-dists_to_centers) * 2.0
def embedding_centers_to_soft_masks_by_distance2(embedding, centers):
"""Masks from embedding and centers based on distance to centers.
Args:
embedding: A tf.float32 tensor of size [N, embedding_size].
centers: A tf.float32 tensor of size [k, embedding_size].
Returns:
A tf.float32 tensor of [k, N] that contains k soft masks.
"""
max_dist = 100.0
dists_to_centers = instance_segmentation_utils.inputs_distances_to_centers(
embedding, centers)
return tf.transpose(
tf.minimum(tf.maximum((max_dist - dists_to_centers) / max_dist, 0.0),
1.0))
def sample_based_on_scores_and_distances(inputs,
scores,
num_samples,
scores_coef):
"""Iteratively samples points based on distances and scores.
Each point i's current total score is computed as
total_score[i] = scores_coef * log(scores[i]) +
log(dist(i, closest previously picked point))
Args:
inputs: A tf.float32 tensor of size [N, dims] containing the input vectors.
scores: A tf.float32 tensor of size [N] containing the scores. Scores
should be all positive.
num_samples: A tf.int32 scalar defining the number of samples.
scores_coef: A tf.float32 coeffcieint that determines that weight assigned
to scores in comparison to distances.
Returns:
input_samples: The input vectors that are sampled. A tf.float32 tensor of
size [num_samples, dims].
indices: A tf.int32 tensor of size [num_samples] containing the indices of
the sampled points.
"""
log_scores = tf.math.log(scores + sys.float_info.min)
dims = tf.shape(inputs)[1]
init_samples = tf.zeros(tf.stack([num_samples, dims]), dtype=tf.float32)
indices = tf.zeros([num_samples], dtype=tf.int32)
index_0 = tf.expand_dims(
tf.cast(tf.math.argmax(scores, axis=0), dtype=tf.int32), axis=0)
init_sample_0 = tf.gather(inputs, index_0)
min_distances = tf.squeeze(
instance_segmentation_utils.inputs_distances_to_centers(
inputs, init_sample_0),
axis=1)
init_samples += tf.pad(
init_sample_0,
paddings=tf.stack([tf.stack([0, num_samples-1]), tf.stack([0, 0])]))
indices += tf.pad(index_0, paddings=[[0, num_samples-1]])
i = tf.constant(1, dtype=tf.int32)
def body(i, init_samples, indices, min_distances):
"""while loop body. Pick the next center given previously picked centers.
Args:
i: For loop index.
init_samples: Initial samples that is modified inside the body.
indices: Indices of picked samples.
min_distances: Minimum distance of the embedding to centers so far.
Returns:
i and init_centers.
"""
best_new_sample_ind = tf.cast(
tf.math.argmax(
scores_coef * log_scores +
tf.math.log(min_distances + sys.float_info.min),
axis=0),
dtype=tf.int32)
indices += tf.pad(tf.stack([best_new_sample_ind]),
paddings=[[i, num_samples-i-1]])
init_samples_i = tf.expand_dims(tf.gather(inputs, best_new_sample_ind), 0)
init_samples += tf.pad(init_samples_i,
paddings=tf.stack([tf.stack([i, num_samples-i-1]),
tf.stack([0, 0])]))
min_distances = tf.minimum(
tf.squeeze(
instance_segmentation_utils.inputs_distances_to_centers(
inputs, init_samples_i),
axis=1), min_distances)
i += 1
return i, init_samples, indices, min_distances
(i, init_samples, indices, min_distances) = tf.while_loop(
lambda i, init_samples, indices, min_distances: i < num_samples,
body,
[i, init_samples, indices, min_distances])
return init_samples, indices