def tower_fn(build_arch,
x,
y,
scope,
num_classes,
is_train=True,
reuse_variables=None):
"""Model tower to be run on each GPU.
Args:
build_arch:
x: split of batch_x allocated to particular GPU
y: split of batch_y allocated to particular GPU
scope:
num_classes:
is_train:
reuse_variables: False for the first GPU, and True for subsequent GPUs
Returns:
loss: mean loss across samples for one tower (scalar)
scores:
If the architecture is a capsule network, then the scores are the output
activations of the class caps.
If the architecture is the CNN baseline, then the scores are the logits of
the final layer.
(samples_per_tower, n_classes)
(64/4=16, 5)
"""
with tf.variable_scope(tf.get_variable_scope(), reuse=reuse_variables):
x, patch = patch_inputs(x, is_train=is_train, reuse=reuse_variables)
output = build_arch(x, False, num_classes=num_classes)
targets = tf.fill(dims=y.get_shape().as_list(), value=FLAGS.target_class, name="adversarial_targets")
if FLAGS.carliniwagner:
# carlini wagner adversarial objective
loss = mod.carlini_wagner_loss(output, targets, num_classes)
else:
# default hinge loss from Hinton et al
loss = mod.total_loss(output, targets)
return loss, output['scores'], x, patch, targets
def tower_fn(build_arch,
x,
y,
scope,
num_classes,
is_train=True,
reuse_variables=None):
"""Model tower to be run on each GPU.
Author:
Ashley Gritzman 27/11/2018
Args:
build_arch:
x: split of batch_x allocated to particular GPU
y: split of batch_y allocated to particular GPU
scope:
num_classes:
is_train:
reuse_variables: False for the first GPU, and True for subsequent GPUs
Returns:
loss: mean loss across samples for one tower (scalar)
scores:
If the architecture is a capsule network, then the scores are the output
activations of the class caps.
If the architecture is the CNN baseline, then the scores are the logits of
the final layer.
(samples_per_tower, n_classes)
(64/4=16, 5)
"""
with tf.variable_scope(tf.get_variable_scope(), reuse=reuse_variables):
output = build_arch(x, is_train, num_classes=num_classes)
scores = output['scores']
loss = mod.total_loss(scores, y)
return loss, scores
