def _match(self, similarity_matrix, num_valid_rows=-1):
"""Bipartite matches a collection rows and columns. A greedy bi-partite.
TODO(rathodv): Add num_valid_columns options to match only that many columns
with all the rows.
Args:
similarity_matrix: Float tensor of shape [N, M] with pairwise similarity
where higher values mean more similar.
num_valid_rows: A scalar or a 1-D tensor with one element describing the
number of valid rows of similarity_matrix to consider for the bipartite
matching. If set to be negative, then all rows from similarity_matrix
are used.
Returns:
match_results: int32 tensor of shape [M] with match_results[i]=-1
meaning that column i is not matched and otherwise that it is matched to
row match_results[i].
"""
# Convert similarity matrix to distance matrix as tf.image.bipartite tries
# to find minimum distance matches.
distance_matrix = -1 * similarity_matrix
_, match_results = image_ops.bipartite_match(
distance_matrix, num_valid_rows)
match_results = tf.reshape(match_results, [-1])
match_results = tf.cast(match_results, tf.int32)
return match_results
def _match(self, similarity_matrix, valid_rows):
"""Bipartite matches a collection rows and columns. A greedy bi-partite.
TODO(rathodv): Add num_valid_columns options to match only that many columns
with all the rows.
Args:
similarity_matrix: Float tensor of shape [N, M] with pairwise similarity
where higher values mean more similar.
valid_rows: A boolean tensor of shape [N] indicating the rows that are
valid.
Returns:
match_results: int32 tensor of shape [M] with match_results[i]=-1
meaning that column i is not matched and otherwise that it is matched to
row match_results[i].
"""
valid_row_sim_matrix = tf.gather(
similarity_matrix, tf.squeeze(tf.where(valid_rows), axis=-1))
invalid_row_sim_matrix = tf.gather(
similarity_matrix,
tf.squeeze(tf.where(tf.logical_not(valid_rows)), axis=-1))
similarity_matrix = tf.concat(
[valid_row_sim_matrix, invalid_row_sim_matrix], axis=0)
# Convert similarity matrix to distance matrix as tf.image.bipartite tries
# to find minimum distance matches.
distance_matrix = -1 * similarity_matrix
num_valid_rows = tf.reduce_sum(tf.to_float(valid_rows))
_, match_results = image_ops.bipartite_match(
distance_matrix, num_valid_rows=num_valid_rows)
match_results = tf.reshape(match_results, [-1])
match_results = tf.cast(match_results, tf.int32)
return match_results
def matching_IoU(Y, Y_hat, epsilon=1e-6):
norm = norm1_tf(Y)
""" j'ai essayer de supprimer les eventuelles categorie avec que des zeros, mais en vain.
(ça serait 2 lignes simple de numpy !!!) """
# indexPos=tf.reshape(tf.where(norm>0),shape=[-1])
# Y=tf.nn.embedding_lookup(tf.transpose(Y,perm=[2,0,1]),indexPos)
# Y=tf.transpose(Y,perm=[1,2,0])
# norm=tf.nn.embedding_lookup(norm,indexPos)
n = Y.get_shape().as_list()[2]
norm_hat = norm1_tf(Y_hat)
Y_ext = tf.expand_dims(Y, 3)
Y_hat_ext = tf.expand_dims(Y_hat, 2)
""" sca[k,l] = sca( Y[:,:,k] Y_hat[:,:,l]) """
sca = norm1_tf(Y_ext * Y_hat_ext)
norm = tf.expand_dims(norm, 1)
norm_hat = tf.expand_dims(norm_hat, 0)
F_IoU = sca / (epsilon + norm + norm_hat)
""" Attention, bipartite_match cherche à minimiser """
r2c, c2r = image_ops.bipartite_match(-F_IoU, n)
A_perm = tf.gather(F_IoU, c2r)
return tf.reduce_sum(tf.diag_part(A_perm))
def _match(self, similarity_matrix, valid_rows):
"""Bipartite matches a collection rows and columns. A greedy bi-partite.
TODO(rathodv): Add num_valid_columns options to match only that many columns
with all the rows.
Args:
similarity_matrix: Float tensor of shape [N, M] with pairwise similarity
where higher values mean more similar.
valid_rows: A boolean tensor of shape [N] indicating the rows that are
valid.
Returns:
match_results: int32 tensor of shape [M] with match_results[i]=-1
meaning that column i is not matched and otherwise that it is matched to
row match_results[i].
"""
valid_row_sim_matrix = tf.gather(similarity_matrix,
tf.squeeze(tf.where(valid_rows), axis=-1))
invalid_row_sim_matrix = tf.gather(
similarity_matrix,
tf.squeeze(tf.where(tf.logical_not(valid_rows)), axis=-1))
similarity_matrix = tf.concat(
[valid_row_sim_matrix, invalid_row_sim_matrix], axis=0)
# Convert similarity matrix to distance matrix as tf.image.bipartite tries
# to find minimum distance matches.
distance_matrix = -1 * similarity_matrix
num_valid_rows = tf.reduce_sum(tf.to_float(valid_rows))
_, match_results = image_ops.bipartite_match(
distance_matrix, num_valid_rows=num_valid_rows)
match_results = tf.reshape(match_results, [-1])
match_results = tf.cast(match_results, tf.int32)
return match_results
def test_bipartite_matching():
tf.InteractiveSession()
A = tf.constant([[-1., -10, 0], [0, 0, -2], [-3, 0, 0]])
r2c, c2r = image_ops.bipartite_match(A, 3)
A_perm = tf.gather(A, c2r)
min_elem = tf.diag_part(A_perm)
print("A\n", A.eval())
print("c2r", c2r.eval())
print("r2c", r2c.eval())
print("A_perm\n", A_perm.eval())
print("min_elem\n", min_elem.eval())
def _BipartiteMatchTest(self, distance_mat, distance_mat_shape,
num_valid_rows,
expected_row_to_col_match,
expected_col_to_row_match):
distance_mat_np = np.array(distance_mat, dtype=np.float32).reshape(
distance_mat_shape)
expected_row_to_col_match_np = np.array(expected_row_to_col_match,
dtype=np.int32)
expected_col_to_row_match_np = np.array(expected_col_to_row_match,
dtype=np.int32)
with self.cached_session():
distance_mat_tf = constant_op.constant(distance_mat_np,
shape=distance_mat_shape)
location_to_prior, prior_to_location = image_ops.bipartite_match(
distance_mat_tf, num_valid_rows)
location_to_prior_np = location_to_prior.eval()
prior_to_location_np = prior_to_location.eval()
self.assertAllEqual(location_to_prior_np, expected_row_to_col_match_np)
self.assertAllEqual(prior_to_location_np, expected_col_to_row_match_np)
def _BipartiteMatchTest(self, distance_mat, distance_mat_shape,
num_valid_rows,
expected_row_to_col_match,
expected_col_to_row_match):
distance_mat_np = np.array(distance_mat, dtype=np.float32).reshape(
distance_mat_shape)
expected_row_to_col_match_np = np.array(expected_row_to_col_match,
dtype=np.int32)
expected_col_to_row_match_np = np.array(expected_col_to_row_match,
dtype=np.int32)
with self.test_session():
distance_mat_tf = constant_op.constant(distance_mat_np,
shape=distance_mat_shape)
location_to_prior, prior_to_location = image_ops.bipartite_match(
distance_mat_tf, num_valid_rows)
location_to_prior_np = location_to_prior.eval()
prior_to_location_np = prior_to_location.eval()
self.assertAllEqual(location_to_prior_np, expected_row_to_col_match_np)
self.assertAllEqual(prior_to_location_np, expected_col_to_row_match_np)
tf.stack([
tf.where(
tf.reduce_max(left_gt_to_anchors_mask, axis=0) > 0,
tf.argmax(left_gt_to_anchors_mask, axis=0), anchors_to_gt),
tf.range(tf.cast(tf.shape(overlap_matrix)[1], tf.int64))
],
axis=1))
return tf.where(
tf.reduce_max(left_gt_to_anchors_mask, axis=0) > 0,
tf.argmax(left_gt_to_anchors_mask, axis=0), match_indices)
print(sess.run([tf.transpose(areas(gt_bboxes), perm=[1, 0])]))
print(sess.run([tf.maximum(tf.expand_dims(t1, 0), tf.expand_dims(t2, 1))]))
print(sess.run([do_dual_max_match(iou_matrix(gt_bboxes, bboxes), 0.65, 0.6)]))
print(sess.run([image_ops.bipartite_match(iou_matrix(gt_bboxes, bboxes), -1)]))
gt_bboxes1 = tf.constant([[0.12, 0.22, 0.33, 0.44], [0.14, 0.2, 0.35, 0.4],
[0.1, 0.32, 0.3, 0.54]])
gt_bboxes2 = tf.constant([[2, 1, 0, 3.]])
print(sess.run(gt_bboxes1 * tf.expand_dims(gt_bboxes2, 0)))
# elems = np.array([[1,2], [2,3], [3,4]])
# alternates = tf.map_fn(lambda x: [x[0][0],x[1][1]], [elems,elems])
# print(sess.run(alternates))
# scores = tf.constant([11,12,13,14,15])
# labels = tf.constant([[111,12,13,14,15], [211,12,13,14,15], [311,12,13,14,15], [411,12,13,14,15],[511,12,13,14,15]])
# num_bboxes = tf.shape(scores)[0]
# left_count = 8 - num_bboxes
