def testCreateOnecloneWithPS(self):
g = tf.Graph()
with g.as_default():
tf.set_random_seed(0)
tf_inputs = tf.constant(self._inputs, dtype=tf.float32)
tf_labels = tf.constant(self._labels, dtype=tf.float32)
model_fn = BatchNormClassifier
model_args = (tf_inputs, tf_labels)
deploy_config = model_deploy.DeploymentConfig(num_clones=1,
num_ps_tasks=1)
self.assertEqual(slim.get_variables(), [])
clones = model_deploy.create_clones(deploy_config, model_fn, model_args)
self.assertEqual(len(slim.get_variables()), 5)
update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)
self.assertEqual(len(update_ops), 2)
optimizer = tf.train.GradientDescentOptimizer(learning_rate=1.0)
total_loss, grads_and_vars = model_deploy.optimize_clones(clones,
optimizer)
self.assertEqual(len(grads_and_vars), len(tf.trainable_variables()))
self.assertEqual(total_loss.op.name, 'total_loss')
for g, v in grads_and_vars:
self.assertDeviceEqual(g.device, '/job:worker')
self.assertDeviceEqual(v.device, '/job:ps/task:0/CPU:0')
def testCreateMulticloneWithPS(self):
g = tf.Graph()
with g.as_default():
tf.set_random_seed(0)
tf_inputs = tf.constant(self._inputs, dtype=tf.float32)
tf_labels = tf.constant(self._labels, dtype=tf.float32)
model_fn = BatchNormClassifier
clone_args = (tf_inputs, tf_labels)
deploy_config = model_deploy.DeploymentConfig(num_clones=2,
num_ps_tasks=2)
self.assertEqual(slim.get_variables(), [])
clones = model_deploy.create_clones(deploy_config, model_fn, clone_args)
self.assertEqual(len(slim.get_variables()), 5)
for i, v in enumerate(slim.get_variables()):
t = i % 2
self.assertDeviceEqual(v.device, '/job:ps/task:%d/device:CPU:0' % t)
self.assertDeviceEqual(v.device, v.value().device)
self.assertEqual(len(clones), 2)
for i, clone in enumerate(clones):
self.assertEqual(
clone.outputs.op.name,
'clone_%d/BatchNormClassifier/fully_connected/Sigmoid' % i)
self.assertEqual(clone.scope, 'clone_%d/' % i)
self.assertDeviceEqual(clone.device, '/job:worker/device:GPU:%d' % i)
def testCreateMulticlone(self):
g = tf.Graph()
with g.as_default():
tf.set_random_seed(0)
tf_inputs = tf.constant(self._inputs, dtype=tf.float32)
tf_labels = tf.constant(self._labels, dtype=tf.float32)
model_fn = BatchNormClassifier
clone_args = (tf_inputs, tf_labels)
num_clones = 4
deploy_config = model_deploy.DeploymentConfig(num_clones=num_clones)
self.assertEqual(slim.get_variables(), [])
clones = model_deploy.create_clones(deploy_config, model_fn, clone_args)
self.assertEqual(len(slim.get_variables()), 5)
for v in slim.get_variables():
self.assertDeviceEqual(v.device, 'CPU:0')
self.assertDeviceEqual(v.value().device, 'CPU:0')
self.assertEqual(len(clones), num_clones)
for i, clone in enumerate(clones):
self.assertEqual(
clone.outputs.op.name,
'clone_%d/BatchNormClassifier/fully_connected/Sigmoid' % i)
update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS, clone.scope)
self.assertEqual(len(update_ops), 2)
self.assertEqual(clone.scope, 'clone_%d/' % i)
self.assertDeviceEqual(clone.device, 'GPU:%d' % i)
def testCreateSingleclone(self):
g = tf.Graph()
with g.as_default():
tf.set_random_seed(0)
tf_inputs = tf.constant(self._inputs, dtype=tf.float32)
tf_labels = tf.constant(self._labels, dtype=tf.float32)
model_fn = BatchNormClassifier
clone_args = (tf_inputs, tf_labels)
deploy_config = model_deploy.DeploymentConfig(num_clones=1)
self.assertEqual(slim.get_variables(), [])
clones = model_deploy.create_clones(deploy_config, model_fn, clone_args)
clone = clones[0]
self.assertEqual(len(slim.get_variables()), 5)
for v in slim.get_variables():
self.assertDeviceEqual(v.device, 'CPU:0')
self.assertDeviceEqual(v.value().device, 'CPU:0')
self.assertEqual(clone.outputs.op.name,
'BatchNormClassifier/fully_connected/Sigmoid')
self.assertEqual(clone.scope, '')
self.assertDeviceEqual(clone.device, '')
self.assertEqual(len(slim.losses.get_losses()), 1)
update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)
self.assertEqual(len(update_ops), 2)
def testCreateOnecloneWithPS(self):
g = tf.Graph()
with g.as_default():
tf.set_random_seed(0)
tf_inputs = tf.constant(self._inputs, dtype=tf.float32)
tf_labels = tf.constant(self._labels, dtype=tf.float32)
model_fn = BatchNormClassifier
clone_args = (tf_inputs, tf_labels)
deploy_config = model_deploy.DeploymentConfig(num_clones=1,
num_ps_tasks=1)
self.assertEqual(slim.get_variables(), [])
clones = model_deploy.create_clones(deploy_config, model_fn, clone_args)
self.assertEqual(len(clones), 1)
clone = clones[0]
self.assertEqual(clone.outputs.op.name,
'BatchNormClassifier/fully_connected/Sigmoid')
self.assertDeviceEqual(clone.device, '/job:worker')
self.assertEqual(clone.scope, '')
self.assertEqual(len(slim.get_variables()), 5)
for v in slim.get_variables():
self.assertDeviceEqual(v.device, '/job:ps/task:0/CPU:0')
self.assertDeviceEqual(v.device, v.value().device)
def main(_):
if not FLAGS.dataset_dir:
raise ValueError(
'You must supply the dataset directory with --dataset_dir')
with tf.Graph().as_default():
######################
# Config model_deploy#
######################
deploy_config = model_deploy.DeploymentConfig(
num_clones=FLAGS.num_clones,
clone_on_cpu=FLAGS.clone_on_cpu,
replica_id=FLAGS.task,
num_replicas=FLAGS.worker_replicas,
num_ps_tasks=FLAGS.num_ps_tasks)
# Create global_step
with tf.device(deploy_config.variables_device()):
global_step = slim.create_global_step()
######################
# Select the dataset #
######################
dataset = dataset_factory.get_dataset(FLAGS.dataset_name,
FLAGS.dataset_split_name,
FLAGS.dataset_dir)
####################
# Select the model #
####################
model_fn = model_factory.get_model(FLAGS.model_name,
num_classes=(dataset.num_classes -
FLAGS.labels_offset),
weight_decay=FLAGS.weight_decay,
is_training=True)
#####################################
# Select the preprocessing function #
#####################################
preprocessing_name = FLAGS.preprocessing_name or FLAGS.model_name
image_preprocessing_fn = preprocessing_factory.get_preprocessing(
preprocessing_name, is_training=True)
##############################################################
# Create a dataset provider that loads data from the dataset #
##############################################################
with tf.device(deploy_config.inputs_device()):
provider = slim.dataset_data_provider.DatasetDataProvider(
dataset,
num_readers=FLAGS.num_readers,
common_queue_capacity=20 * FLAGS.batch_size,
common_queue_min=10 * FLAGS.batch_size)
[image, label] = provider.get(['image', 'label'])
label -= FLAGS.labels_offset
if FLAGS.train_image_size is None:
train_image_size = model_fn.default_image_size
else:
train_image_size = FLAGS.train_image_size
image = image_preprocessing_fn(image, train_image_size,
train_image_size)
images, labels = tf.train.batch(
[image, label],
batch_size=FLAGS.batch_size,
num_threads=FLAGS.num_preprocessing_threads,
capacity=5 * FLAGS.batch_size)
labels = slim.one_hot_encoding(
labels, dataset.num_classes - FLAGS.labels_offset)
batch_queue = slim.prefetch_queue.prefetch_queue(
[images, labels], capacity=2 * deploy_config.num_clones)
####################
# Define the model #
####################
def clone_fn(batch_queue):
"""Allows data parallelism by creating multiple clones of the model_fn."""
images, labels = batch_queue.dequeue()
logits, end_points = model_fn(images)
#############################
# Specify the loss function #
#############################
if 'AuxLogits' in end_points:
slim.losses.softmax_cross_entropy(
end_points['AuxLogits'],
labels,
label_smoothing=FLAGS.label_smoothing,
weight=0.4,
scope='aux_loss')
slim.losses.softmax_cross_entropy(
logits,
labels,
label_smoothing=FLAGS.label_smoothing,
weight=1.0)
# Gather initial summaries.
summaries = set(tf.get_collection(tf.GraphKeys.SUMMARIES))
clones = model_deploy.create_clones(deploy_config, clone_fn,
[batch_queue])
first_clone_scope = deploy_config.clone_scope(0)
# Gather update_ops from the first clone. These contain, for example,
# the updates for the batch_norm variables created by model_fn.
update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS,
first_clone_scope)
# Add summaries for losses.
for loss in tf.get_collection(tf.GraphKeys.LOSSES, first_clone_scope):
tf.scalar_summary('losses/%s' % loss.op.name, loss)
# Add summaries for variables.
for variable in slim.get_model_variables():
summaries.add(tf.histogram_summary(variable.op.name, variable))
#################################
# Configure the moving averages #
#################################
if FLAGS.moving_average_decay:
moving_average_variables = slim.get_model_variables()
variable_averages = tf.train.ExponentialMovingAverage(
FLAGS.moving_average_decay, global_step)
else:
moving_average_variables, variable_averages = None, None
#########################################
# Configure the optimization procedure. #
#########################################
with tf.device(deploy_config.optimizer_device()):
learning_rate = _configure_learning_rate(dataset.num_samples,
global_step)
optimizer = _configure_optimizer(learning_rate)
summaries.add(
tf.scalar_summary('learning_rate',
learning_rate,
name='learning_rate'))
if FLAGS.sync_replicas:
# If sync_replicas is enabled, the averaging will be done in the chief
# queue runner.
optimizer = tf.train.SyncReplicasOptimizer(
opt=optimizer,
replicas_to_aggregate=FLAGS.replicas_to_aggregate,
variable_averages=variable_averages,
variables_to_average=moving_average_variables,
replica_id=tf.constant(FLAGS.task, tf.int32, shape=()),
total_num_replicas=FLAGS.worker_replicas)
elif FLAGS.moving_average_decay:
# Update ops executed locally by trainer.
update_ops.append(
variable_averages.apply(moving_average_variables))
# TODO(sguada) Refactor into function that takes the clones and optimizer
# and returns a train_tensor and summary_op
total_loss, clones_gradients = model_deploy.optimize_clones(
clones, optimizer)
# Add total_loss to summary.
summaries.add(
tf.scalar_summary('total_loss', total_loss, name='total_loss'))
# Create gradient updates.
grad_updates = optimizer.apply_gradients(clones_gradients,
global_step=global_step)
update_ops.append(grad_updates)
update_op = tf.group(*update_ops)
train_tensor = control_flow_ops.with_dependencies([update_op],
total_loss,
name='train_op')
# Add the summaries from the first clone. These contain the summaries
# created by model_fn and either optimize_clones() or _gather_clone_loss().
summaries |= set(
tf.get_collection(tf.GraphKeys.SUMMARIES, first_clone_scope))
# Merge all summaries together.
summary_op = tf.merge_summary(list(summaries), name='summary_op')
###########################
# Kicks off the training. #
###########################
slim.learning.train(
train_tensor,
logdir=FLAGS.train_dir,
master=FLAGS.master,
is_chief=(FLAGS.task == 0),
init_fn=_get_init_fn(),
summary_op=summary_op,
number_of_steps=FLAGS.max_number_of_steps,
log_every_n_steps=FLAGS.log_every_n_steps,
save_summaries_secs=FLAGS.save_summaries_secs,
save_interval_secs=FLAGS.save_interval_secs,
sync_optimizer=optimizer if FLAGS.sync_replicas else None)
