def batch_center_loss(self,
u,
label_u,
loss_direction,
loss_scale,
knn_k,
normed=False):
if normed:
per_img_avg = tfdist.normed_euclidean2(u, tf.reduce_mean(u, 0))
else:
per_img_avg = tfdist.euclidean(u, tf.reduce_mean(u, 0))
if loss_scale == 'similarity':
per_img_avg_in_scale = tf.negative(
tf.exp(tf.cast(tf.negative(per_img_avg), dtype=tf.float32)))
else:
per_img_avg_in_scale = per_img_avg
loss_before_direction = tf.reduce_mean(per_img_avg_in_scale)
if loss_direction == 'increase':
loss = tf.negative(loss_before_direction)
else:
loss = loss_before_direction
return loss
def knn_loss(self,
u,
label_u,
loss_direction,
loss_scale,
knn_k,
normed=False):
if knn_k is not None:
k = knn_k
distances = tfdist.distance(u, u, pair=True, dist_type='euclidean')
values, indices = tf.math.top_k(-distances, k=k, sorted=True)
if normed:
per_img_avg = tfdist.normed_euclidean2(
u, tf.reduce_mean(tf.gather(u, indices), 1))
else:
per_img_avg = tfdist.euclidean(
u, tf.reduce_mean(tf.gather(u, indices), 1))
if loss_scale == 'similarity':
per_img_avg_in_scale = tf.negative(
tf.exp(tf.cast(tf.negative(per_img_avg), dtype=tf.float32)))
else:
per_img_avg_in_scale = per_img_avg
loss_before_direction = tf.reduce_mean(per_img_avg_in_scale)
if loss_direction == 'increase':
loss = tf.negative(loss_before_direction)
else:
loss = loss_before_direction
return loss
def class_knn_loss(self,
u,
label_u,
loss_direction,
loss_scale,
knn_k,
normed=False):
if knn_k is not None:
k = knn_k
indices = tf.constant(0, shape=[0, k], dtype=tf.int32)
iters = tf.cond(tf.equal(self.stage, 0), lambda: self.batch_size,
lambda: self.val_batch_size)
def condition(indices, i):
return tf.less(i, iters)
def body(indices, i):
distances = tfdist.distance(u, u, pair=True, dist_type='euclidean')
b1_nt = tf.where(tf.equal(tf.gather(label_u, i, axis=0), 1))
corrected_distances_b1 = tf.where(
tf.squeeze(
tf.reduce_any(tf.equal(tf.gather(label_u, b1_nt, axis=1),
1),
axis=1)), tf.gather(distances, i, axis=0),
tf.multiply(tf.ones_like(tf.gather(distances, i, axis=0)),
float("inf")))
val, ind = tf.math.top_k(-corrected_distances_b1, k=k, sorted=True)
return [
tf.concat([indices, tf.reshape(ind, shape=[1, -1])], 0),
tf.add(i, 1)
]
indices, i = tf.while_loop(condition,
body, [indices, 0],
shape_invariants=[
tf.TensorShape([None, None]),
tf.TensorShape([])
])
if normed:
per_img_avg = tfdist.normed_euclidean2(
u, tf.reduce_mean(tf.gather(u, indices), 1))
else:
per_img_avg = tfdist.euclidean(
u, tf.reduce_mean(tf.gather(u, indices), 1))
if loss_scale == 'similarity':
per_img_avg_in_scale = tf.negative(
tf.exp(tf.cast(tf.negative(per_img_avg), dtype=tf.float32)))
else:
per_img_avg_in_scale = per_img_avg
loss_before_direction = tf.reduce_mean(per_img_avg_in_scale)
if loss_direction == 'increase':
loss = tf.negative(loss_before_direction)
else:
loss = loss_before_direction
return loss
def class_center_loss(self,
u,
label_u,
loss_direction,
loss_scale,
knn_k,
normed=False):
if self.extract_features or self.extract_hashlayer_features:
shape1 = label_u.shape[1].value
targets = tf.constant(0.0, shape=[0, u.shape[1]])
for i in range(0, shape1):
targets = tf.concat([
targets,
tf.stop_gradient(
tf.reshape(
tf.reduce_mean(
tf.reshape(
tf.gather(
u, tf.where(tf.equal(label_u[:, i],
1))),
[-1, u.shape[1]]), 0), [1, -1]))
], 0)
corrected_targets = tf.where(tf.is_nan(targets),
tf.zeros_like(targets), targets)
targets = tf.stop_gradient(corrected_targets)
else:
targets = self.targets
mean = tf.divide(
tf.reduce_sum(
tf.multiply(
tf.cast(tf.multiply(tf.expand_dims(label_u, 2),
np.ones((1, 1, np.int(u.shape[1])))),
dtype=tf.float32), targets), 1),
tf.reshape(tf.cast(tf.reduce_sum(label_u, 1), dtype=tf.float32),
(-1, 1)))
if normed:
per_img_avg = tfdist.normed_euclidean2(u, mean)
else:
per_img_avg = tfdist.euclidean(u, mean)
if loss_scale == 'similarity':
per_img_avg_in_scale = tf.negative(
tf.exp(tf.cast(tf.negative(per_img_avg), dtype=tf.float32)))
else:
per_img_avg_in_scale = per_img_avg
loss_before_direction = tf.reduce_mean(per_img_avg_in_scale)
if loss_direction == 'increase':
loss = tf.negative(loss_before_direction)
else:
loss = loss_before_direction
return loss