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/device:GPU:0')
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 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/device:GPU:0')
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 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 testCreateLogisticClassifier(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 = LogisticClassifier
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()), 2)
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,
'LogisticClassifier/fully_connected/Sigmoid')
self.assertEqual(clone.scope, '')
self.assertDeviceEqual(clone.device, 'GPU:0')
self.assertEqual(len(slim.losses.get_losses()), 1)
update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)
self.assertEqual(update_ops, [])
def testCPUonly(self):
deploy_config = model_deploy.DeploymentConfig(clone_on_cpu=True)
self.assertEqual(deploy_config.caching_device(), None)
self.assertDeviceEqual(deploy_config.clone_device(0), 'CPU:0')
self.assertEqual(deploy_config.clone_scope(0), '')
self.assertDeviceEqual(deploy_config.optimizer_device(), 'CPU:0')
self.assertDeviceEqual(deploy_config.inputs_device(), 'CPU:0')
self.assertDeviceEqual(deploy_config.variables_device(), 'CPU:0')
def testDefaults(self):
deploy_config = model_deploy.DeploymentConfig()
self.assertEqual(slim.get_variables(), [])
self.assertEqual(deploy_config.caching_device(), None)
self.assertDeviceEqual(deploy_config.clone_device(0), 'GPU:0')
self.assertEqual(deploy_config.clone_scope(0), '')
self.assertDeviceEqual(deploy_config.optimizer_device(), 'CPU:0')
self.assertDeviceEqual(deploy_config.inputs_device(), 'CPU:0')
self.assertDeviceEqual(deploy_config.variables_device(), 'CPU:0')
def testMultiGPU(self):
deploy_config = model_deploy.DeploymentConfig(num_clones=2)
self.assertEqual(deploy_config.caching_device(), None)
self.assertDeviceEqual(deploy_config.clone_device(0), 'GPU:0')
self.assertDeviceEqual(deploy_config.clone_device(1), 'GPU:1')
self.assertEqual(deploy_config.clone_scope(0), 'clone_0')
self.assertEqual(deploy_config.clone_scope(1), 'clone_1')
self.assertDeviceEqual(deploy_config.optimizer_device(), 'CPU:0')
self.assertDeviceEqual(deploy_config.inputs_device(), 'CPU:0')
self.assertDeviceEqual(deploy_config.variables_device(), 'CPU:0')
def testReplicasPS(self):
deploy_config = model_deploy.DeploymentConfig(num_replicas=2,
num_ps_tasks=2)
self.assertDeviceEqual(deploy_config.clone_device(0),
'/job:worker/device:GPU:0')
self.assertEqual(deploy_config.clone_scope(0), '')
self.assertDeviceEqual(deploy_config.optimizer_device(),
'/job:worker/device:CPU:0')
self.assertDeviceEqual(deploy_config.inputs_device(),
'/job:worker/device:CPU:0')
def testLocalTrainOp(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=2,
clone_on_cpu=True)
optimizer = tf.train.GradientDescentOptimizer(learning_rate=1.0)
self.assertEqual(slim.get_variables(), [])
model = model_deploy.deploy(deploy_config,
model_fn,
model_args,
optimizer=optimizer)
update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)
self.assertEqual(len(update_ops), 4)
self.assertEqual(len(model.clones), 2)
self.assertEqual(model.total_loss.op.name, 'total_loss')
self.assertEqual(model.summary_op.op.name, 'summary_op/summary_op')
self.assertEqual(model.train_op.op.name, 'train_op')
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
moving_mean = tf.contrib.framework.get_variables_by_name(
'moving_mean')[0]
moving_variance = tf.contrib.framework.get_variables_by_name(
'moving_variance')[0]
initial_loss = sess.run(model.total_loss)
initial_mean, initial_variance = sess.run(
[moving_mean, moving_variance])
self.assertAllClose(initial_mean, [0.0, 0.0, 0.0, 0.0])
self.assertAllClose(initial_variance, [1.0, 1.0, 1.0, 1.0])
for _ in range(10):
sess.run(model.train_op)
final_loss = sess.run(model.total_loss)
self.assertLess(final_loss, initial_loss / 5.0)
final_mean, final_variance = sess.run(
[moving_mean, moving_variance])
expected_mean = np.array([0.125, 0.25, 0.375, 0.25])
expected_var = np.array([0.109375, 0.1875, 0.234375, 0.1875])
expected_var = self._addBesselsCorrection(16, expected_var)
self.assertAllClose(final_mean, expected_mean)
self.assertAllClose(final_variance, expected_var)
def testMultiGPUPS(self):
deploy_config = model_deploy.DeploymentConfig(num_clones=2,
num_ps_tasks=1)
self.assertEqual(deploy_config.caching_device()(tf.no_op()), '')
self.assertDeviceEqual(deploy_config.clone_device(0),
'/job:worker/device:GPU:0')
self.assertDeviceEqual(deploy_config.clone_device(1),
'/job:worker/device:GPU:1')
self.assertEqual(deploy_config.clone_scope(0), 'clone_0')
self.assertEqual(deploy_config.clone_scope(1), 'clone_1')
self.assertDeviceEqual(deploy_config.optimizer_device(),
'/job:worker/device:CPU:0')
self.assertDeviceEqual(deploy_config.inputs_device(),
'/job:worker/device:CPU:0')
def testVariablesPS(self):
deploy_config = model_deploy.DeploymentConfig(num_ps_tasks=2)
with tf.device(deploy_config.variables_device()):
a = tf.Variable(0)
b = tf.Variable(0)
c = tf.no_op()
d = slim.variable('a', [],
caching_device=deploy_config.caching_device())
self.assertDeviceEqual(a.device, '/job:ps/task:0/device:CPU:0')
self.assertDeviceEqual(a.device, a.value().device)
self.assertDeviceEqual(b.device, '/job:ps/task:1/device:CPU:0')
self.assertDeviceEqual(b.device, b.value().device)
self.assertDeviceEqual(c.device, '')
self.assertDeviceEqual(d.device, '/job:ps/task:0/device:CPU:0')
self.assertDeviceEqual(d.value().device, '')
def testNoSummariesOnGPUForEvals(self):
with tf.Graph().as_default():
deploy_config = model_deploy.DeploymentConfig(num_clones=2)
# clone function creates a fully_connected layer with a regularizer loss.
def ModelFn():
inputs = tf.constant(1.0, shape=(10, 20), dtype=tf.float32)
reg = tf.contrib.layers.l2_regularizer(0.001)
tf.contrib.layers.fully_connected(inputs,
30,
weights_regularizer=reg)
# No optimizer here, it's an eval.
model = model_deploy.deploy(deploy_config, ModelFn)
# The model summary op should have a few summary inputs and all of them
# should be on the CPU.
self.assertTrue(model.summary_op.op.inputs)
for inp in model.summary_op.op.inputs:
self.assertEqual('/device:CPU:0', inp.device)
def testPS(self):
deploy_config = model_deploy.DeploymentConfig(num_clones=1,
num_ps_tasks=1)
self.assertDeviceEqual(deploy_config.clone_device(0),
'/job:worker/device:GPU:0')
self.assertEqual(deploy_config.clone_scope(0), '')
self.assertDeviceEqual(deploy_config.optimizer_device(),
'/job:worker/device:CPU:0')
self.assertDeviceEqual(deploy_config.inputs_device(),
'/job:worker/device:CPU:0')
with tf.device(deploy_config.variables_device()):
a = tf.Variable(0)
b = tf.Variable(0)
c = tf.no_op()
d = slim.variable('a', [],
caching_device=deploy_config.caching_device())
self.assertDeviceEqual(a.device, '/job:ps/task:0/device:CPU:0')
self.assertDeviceEqual(a.device, a.value().device)
self.assertDeviceEqual(b.device, '/job:ps/task:0/device:CPU:0')
self.assertDeviceEqual(b.device, b.value().device)
self.assertDeviceEqual(c.device, '')
self.assertDeviceEqual(d.device, '/job:ps/task:0/device:CPU:0')
self.assertDeviceEqual(d.value().device, '')
def main(_):
if not FLAGS.dataset_dir:
raise ValueError(
'You must supply the dataset directory with --dataset_dir')
tf.logging.set_verbosity(tf.logging.INFO)
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 network #
######################
network_fn = nets_factory.get_network_fn(
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
train_image_size = FLAGS.train_image_size or network_fn.default_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 network_fn."""
images, labels = batch_queue.dequeue()
logits, end_points = network_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,
weights=0.4,
scope='aux_loss')
slim.losses.softmax_cross_entropy(
logits,
labels,
label_smoothing=FLAGS.label_smoothing,
weights=1.0)
return end_points
# 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 network_fn.
update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS,
first_clone_scope)
# Add summaries for end_points.
end_points = clones[0].outputs
for end_point in end_points:
x = end_points[end_point]
summaries.add(tf.summary.histogram('activations/' + end_point, x))
summaries.add(
tf.summary.scalar('sparsity/' + end_point,
tf.nn.zero_fraction(x)))
# Add summaries for losses.
for loss in tf.get_collection(tf.GraphKeys.LOSSES, first_clone_scope):
summaries.add(tf.summary.scalar('losses/%s' % loss.op.name, loss))
# Add summaries for variables.
for variable in slim.get_model_variables():
summaries.add(tf.summary.histogram(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.summary.scalar('learning_rate', 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,
total_num_replicas=FLAGS.worker_replicas,
variable_averages=variable_averages,
variables_to_average=moving_average_variables)
elif FLAGS.moving_average_decay:
# Update ops executed locally by trainer.
update_ops.append(
variable_averages.apply(moving_average_variables))
# Variables to train.
variables_to_train = _get_variables_to_train()
# and returns a train_tensor and summary_op
total_loss, clones_gradients = model_deploy.optimize_clones(
clones, optimizer, var_list=variables_to_train)
# Add total_loss to summary.
summaries.add(tf.summary.scalar('total_loss', 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)
with tf.control_dependencies([update_op]):
train_tensor = tf.identity(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.summary.merge(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)