def perception_loss(self, images, labels, num_classes, is_training, restore, scope):
"""Calculate the total loss on a single tower running the ImageNet model.
We perform 'batch splitting'. This means that we cut up a batch across
multiple GPU's. For instance, if the batch size = 32 and num_gpus = 2,
then each tower will operate on an batch of 16 images.
Args:
images: Images. 4D tensor of size [batch_size, FLAGS.image_size,
FLAGS.image_size, 3].
labels: 1-D integer Tensor of [batch_size].
num_classes: number of classes
scope: unique prefix string identifying the ImageNet tower, e.g.
'tower_0'.
Returns:
Tensor of shape [] containing the total loss for a batch of data
"""
# When fine-tuning a model, we do not restore the logits but instead we
# randomly initialize the logits. The number of classes in the output of the
# logit is the number of classes in specified Dataset.
# Build inference Graph.
with tf.name_scope(scope) as scope:
with tf.device('/gpu:0'):
logits = inception.inference(tf.tile(images,[1,1,1,3]), num_classes, for_training=is_training, restore_logits=restore, scope=scope)
# Build the portion of the Graph calculating the losses. Note that we will
# assemble the total_loss using a custom function below.
split_batch_size = images.get_shape().as_list()[0]
inception.loss(logits, labels, batch_size=split_batch_size)
# Assemble all of the losses for the current tower only.
losses = tf.get_collection(slim.losses.LOSSES_COLLECTION, scope)
# Calculate the total loss for the current tower.
total_loss = tf.add_n(losses, name='total_loss')
# Attach a scalar summmary to all individual losses and the total loss; do the
# same for the averaged version of the losses.
for l in losses:
# Remove 'tower_[0-9]/' from the name in case this is a multi-GPU training
# session. This helps the clarity of presentation on TensorBoard.
loss_name = l.op.name
# Name each loss as '(raw)' and name the moving average version of the loss
# as the original loss name.
tf.scalar_summary(loss_name, l)
return total_loss
def _tower_loss(images, labels, num_classes, scope, reuse_variables=None):
"""Calculate the total loss on a single tower running the ImageNet model.
We perform 'batch splitting'. This means that we cut up a batch across
multiple GPUs. For instance, if the batch size = 32 and num_gpus = 2,
then each tower will operate on an batch of 16 images.
Args:
images: Images. 4D tensor of size [batch_size, FLAGS.image_size,
FLAGS.image_size, 3].
labels: 1-D integer Tensor of [batch_size].
num_classes: number of classes
scope: unique prefix string identifying the ImageNet tower, e.g.
'tower_0'.
Returns:
Tensor of shape [] containing the total loss for a batch of data
"""
# When fine-tuning a model, we do not restore the logits but instead we
# randomly initialize the logits. The number of classes in the output of the
# logit is the number of classes in specified Dataset.
restore_logits = not FLAGS.fine_tune
# Build inference Graph.
with tf.variable_scope(tf.get_variable_scope(), reuse=reuse_variables):
logits = inception.inference(images,
num_classes,
for_training=True,
restore_logits=restore_logits,
scope=scope)
# Build the portion of the Graph calculating the losses. Note that we will
# assemble the total_loss using a custom function below.
split_batch_size = images.get_shape().as_list()[0]
inception.loss(logits, labels, batch_size=split_batch_size)
# Assemble all of the losses for the current tower only.
losses = tf.get_collection(slim.losses.LOSSES_COLLECTION, scope)
# Calculate the total loss for the current tower.
regularization_losses = tf.get_collection(
tf.GraphKeys.REGULARIZATION_LOSSES)
total_loss = tf.add_n(losses + regularization_losses, name='total_loss')
return total_loss
def main(argv=None):
ps_hosts = FLAGS.ps_hosts.split(',')
worker_hosts = FLAGS.worker_hosts.split(',')
tf.logging.info('PS hosts are: %s' % ps_hosts)
tf.logging.info('Worker hosts are: %s' % worker_hosts)
cluster_spec = tf.train.ClusterSpec({
'ps': ps_hosts,
'worker': worker_hosts
})
server = tf.train.Server({
'ps': ps_hosts,
'worker': worker_hosts
},
job_name=FLAGS.job_name,
task_index=FLAGS.task_id,
protocol=FLAGS.protocol)
sspManager = SspManager(len(worker_hosts), 5)
if FLAGS.job_name == 'ps':
if FLAGS.task_id == 0:
rpcServer = sspManager.create_rpc_server(ps_hosts[0].split(':')[0])
rpcServer.serve()
server.join()
time.sleep(5)
rpcClient = sspManager.create_rpc_client(ps_hosts[0].split(':')[0])
dataset = ImagenetData(subset=FLAGS.subset)
assert dataset.data_files()
is_chief = (FLAGS.task_id == 0)
if is_chief:
if not tf.gfile.Exists(FLAGS.train_dir):
tf.gfile.MakeDirs(FLAGS.train_dir)
num_workers = len(cluster_spec.as_dict()['worker'])
num_parameter_servers = len(cluster_spec.as_dict()['ps'])
with tf.device('/job:worker/task:%d' % FLAGS.task_id):
with slim.scopes.arg_scope(
[slim.variables.variable, slim.variables.global_step],
device=slim.variables.VariableDeviceChooser(
num_parameter_servers)):
'''Prepare Input'''
global_step = slim.variables.global_step()
batch_size = tf.placeholder(dtype=tf.int32,
shape=(),
name='batch_size')
images, labels = image_processing.distorted_inputs(
dataset,
batch_size,
num_preprocess_threads=FLAGS.num_preprocess_threads)
num_classes = dataset.num_classes() + 1
'''Inference'''
logits = inception.inference(images,
num_classes,
for_training=True)
'''Loss'''
inception.loss(logits, labels, batch_size)
losses = tf.get_collection(slim.losses.LOSSES_COLLECTION)
losses += tf.get_collection(tf.GraphKeys.REGULARIZATION_LOSSES)
total_loss = tf.add_n(losses, name='total_loss')
if is_chief:
loss_averages = tf.train.ExponentialMovingAverage(0.9,
name='avg')
loss_averages_op = loss_averages.apply(losses + [total_loss])
with tf.control_dependencies([loss_averages_op]):
total_loss = tf.identity(total_loss)
'''Optimizer'''
exp_moving_averager = tf.train.ExponentialMovingAverage(
inception.MOVING_AVERAGE_DECAY, global_step)
variables_to_average = (tf.trainable_variables() +
tf.moving_average_variables())
num_batches_per_epoch = (dataset.num_examples_per_epoch() /
FLAGS.batch_size)
decay_steps = int(num_batches_per_epoch *
FLAGS.num_epochs_per_decay / num_workers)
lr = tf.train.exponential_decay(FLAGS.initial_learning_rate,
global_step,
decay_steps,
FLAGS.learning_rate_decay_factor,
staircase=True)
opt = tf.train.RMSPropOptimizer(lr,
RMSPROP_DECAY,
momentum=RMSPROP_MOMENTUM,
epsilon=RMSPROP_EPSILON)
'''Train Operation'''
batchnorm_updates = tf.get_collection(
slim.ops.UPDATE_OPS_COLLECTION)
assert batchnorm_updates, 'Batchnorm updates are missing'
batchnorm_updates_op = tf.group(*batchnorm_updates)
with tf.control_dependencies([batchnorm_updates_op]):
total_loss = tf.identity(total_loss)
naive_grads = opt.compute_gradients(total_loss)
grads = [(tf.scalar_mul(
tf.cast(batch_size / FLAGS.batch_size, tf.float32), grad), var)
for grad, var in naive_grads]
apply_gradients_op = opt.apply_gradients(grads,
global_step=global_step)
with tf.control_dependencies([apply_gradients_op]):
train_op = tf.identity(total_loss, name='train_op')
'''Supervisor and Session'''
saver = tf.train.Saver()
init_op = tf.global_variables_initializer()
sv = tf.train.Supervisor(is_chief=is_chief,
logdir=FLAGS.train_dir,
init_op=init_op,
summary_op=None,
global_step=global_step,
recovery_wait_secs=1,
saver=saver,
save_model_secs=FLAGS.save_interval_secs)
tf.logging.info('%s Supervisor' % datetime.now())
sess_config = tf.ConfigProto(
allow_soft_placement=True,
log_device_placement=FLAGS.log_device_placement)
sess = sv.prepare_or_wait_for_session(server.target,
config=sess_config)
queue_runners = tf.get_collection(tf.GraphKeys.QUEUE_RUNNERS)
'''Start Training'''
sv.start_queue_runners(sess, queue_runners)
tf.logging.info('Started %d queues for processing input data.',
len(queue_runners))
batch_size_num = FLAGS.batch_size
for step in range(FLAGS.max_steps):
start_time = time.time()
run_options = tf.RunOptions(
trace_level=tf.RunOptions.FULL_TRACE)
run_metadata = tf.RunMetadata()
loss_value, gs = sess.run(
[train_op, global_step],
feed_dict={batch_size: batch_size_num},
options=run_options,
run_metadata=run_metadata)
assert not np.isnan(
loss_value), 'Model diverged with loss = NaN'
duration = time.time() - start_time
examples_per_sec = batch_size_num / float(duration)
sec_per_batch = float(duration)
format_str = (
"time: " + str(time.time()) +
'; %s: step %d (gs %d), loss= %.2f (%.1f samples/s; %.3f s/batch)'
)
tf.logging.info(format_str %
(datetime.now(), step, gs, loss_value,
examples_per_sec, sec_per_batch))
rpcClient.check_staleness(FLAGS.task_id, step)
def _tower_loss(images, labels, num_classes, scope, reuse_variables=None):
"""Calculate the total loss on a single tower running the ImageNet model.
We perform 'batch splitting'. This means that we cut up a batch across
multiple GPU's. For instance, if the batch size = 32 and num_gpus = 2,
then each tower will operate on an batch of 16 images.
Args:
images: Images. 4D tensor of size [batch_size, FLAGS.image_size,
FLAGS.image_size, 3].
labels: 1-D integer Tensor of [batch_size].
num_classes: number of classes
scope: unique prefix string identifying the ImageNet tower, e.g.
'tower_0'.
Returns:
Tensor of shape [] containing the total loss for a batch of data
"""
# When fine-tuning a model, we do not restore the logits but instead we
# randomly initialize the logits. The number of classes in the output of the
# logit is the number of classes in specified Dataset.
restore_logits = not FLAGS.fine_tune
# Build inference Graph.
with tf.variable_scope(tf.get_variable_scope(), reuse=reuse_variables):
logits = inception.inference(images,
num_classes,
for_training=True,
restore_logits=restore_logits,
scope=scope)
# Build the portion of the Graph calculating the losses. Note that we will
# assemble the total_loss using a custom function below.
split_batch_size = images.get_shape().as_list()[0]
inception.loss(logits, labels, batch_size=split_batch_size)
# Assemble all of the losses for the current tower only.
losses = tf.get_collection(slim.losses.LOSSES_COLLECTION, scope)
# Calculate the total loss for the current tower.
regularization_losses = tf.get_collection(
tf.GraphKeys.REGULARIZATION_LOSSES)
total_loss = tf.add_n(losses + regularization_losses, name='total_loss')
# Compute the moving average of all individual losses and the total loss.
loss_averages = tf.train.ExponentialMovingAverage(0.9, name='avg')
loss_averages_op = loss_averages.apply(losses + [total_loss])
# Attach a scalar summmary to all individual losses and the total loss; do the
# same for the averaged version of the losses.
for l in losses + [total_loss]:
# Remove 'tower_[0-9]/' from the name in case this is a multi-GPU training
# session. This helps the clarity of presentation on TensorBoard.
loss_name = re.sub('%s_[0-9]*/' % inception.TOWER_NAME, '', l.op.name)
# Name each loss as '(raw)' and name the moving average version of the loss
# as the original loss name.
tf.summary.scalar(loss_name + ' (raw)', l)
tf.summary.scalar(loss_name, loss_averages.average(l))
with tf.control_dependencies([loss_averages_op]):
total_loss = tf.identity(total_loss)
return total_loss
def train(target, dataset, cluster_spec):
"""Train Inception on a dataset for a number of steps."""
# Number of workers and parameter servers are infered from the workers and ps
# hosts string.
num_workers = len(cluster_spec.as_dict()['worker'])
num_parameter_servers = len(cluster_spec.as_dict()['ps'])
# If no value is given, num_replicas_to_aggregate defaults to be the number of
# workers.
if FLAGS.num_replicas_to_aggregate == -1:
num_replicas_to_aggregate = num_workers
else:
num_replicas_to_aggregate = FLAGS.num_replicas_to_aggregate
# Both should be greater than 0 in a distributed training.
assert num_workers > 0 and num_parameter_servers > 0, (' num_workers and '
'num_parameter_servers'
' must be > 0.')
# Choose worker 0 as the chief. Note that any worker could be the chief
# but there should be only one chief.
is_chief = (FLAGS.task_id == 0)
# Ops are assigned to worker by default.
with tf.device('/job:worker/task:%d' % FLAGS.task_id):
# Variables and its related init/assign ops are assigned to ps.
with slim.scopes.arg_scope(
[slim.variables.variable, slim.variables.global_step],
device=slim.variables.VariableDeviceChooser(num_parameter_servers)):
# Create a variable to count the number of train() calls. This equals the
# number of updates applied to the variables.
global_step = slim.variables.global_step()
# Calculate the learning rate schedule.
num_batches_per_epoch = (dataset.num_examples_per_epoch() /
FLAGS.batch_size)
# Decay steps need to be divided by the number of replicas to aggregate.
decay_steps = int(num_batches_per_epoch * FLAGS.num_epochs_per_decay /
num_replicas_to_aggregate)
# Decay the learning rate exponentially based on the number of steps.
lr = tf.train.exponential_decay(FLAGS.initial_learning_rate,
global_step,
decay_steps,
FLAGS.learning_rate_decay_factor,
staircase=True)
# Add a summary to track the learning rate.
tf.summary.scalar('learning_rate', lr)
# Create an optimizer that performs gradient descent.
opt = tf.train.RMSPropOptimizer(lr,
RMSPROP_DECAY,
momentum=RMSPROP_MOMENTUM,
epsilon=RMSPROP_EPSILON)
images, labels = image_processing.distorted_inputs(
dataset,
batch_size=FLAGS.batch_size,
num_preprocess_threads=FLAGS.num_preprocess_threads)
# Number of classes in the Dataset label set plus 1.
# Label 0 is reserved for an (unused) background class.
num_classes = dataset.num_classes() + 1
logits = inception.inference(images, num_classes, for_training=True)
# Add classification loss.
inception.loss(logits, labels)
# Gather all of the losses including regularization losses.
losses = tf.get_collection(slim.losses.LOSSES_COLLECTION)
losses += tf.get_collection(tf.GraphKeys.REGULARIZATION_LOSSES)
total_loss = tf.add_n(losses, name='total_loss')
if is_chief:
# Compute the moving average of all individual losses and the
# total loss.
loss_averages = tf.train.ExponentialMovingAverage(0.9, name='avg')
loss_averages_op = loss_averages.apply(losses + [total_loss])
# Attach a scalar summmary to all individual losses and the total loss;
# do the same for the averaged version of the losses.
for l in losses + [total_loss]:
loss_name = l.op.name
# Name each loss as '(raw)' and name the moving average version of the
# loss as the original loss name.
tf.summary.scalar(loss_name + ' (raw)', l)
tf.summary.scalar(loss_name, loss_averages.average(l))
# Add dependency to compute loss_averages.
with tf.control_dependencies([loss_averages_op]):
total_loss = tf.identity(total_loss)
# Track the moving averages of all trainable variables.
# Note that we maintain a 'double-average' of the BatchNormalization
# global statistics.
# This is not needed when the number of replicas are small but important
# for synchronous distributed training with tens of workers/replicas.
exp_moving_averager = tf.train.ExponentialMovingAverage(
inception.MOVING_AVERAGE_DECAY, global_step)
variables_to_average = (
tf.trainable_variables() + tf.moving_average_variables())
# Add histograms for model variables.
for var in variables_to_average:
tf.summary.histogram(var.op.name, var)
# Create synchronous replica optimizer.
opt = tf.train.SyncReplicasOptimizer(
opt,
replicas_to_aggregate=num_replicas_to_aggregate,
replica_id=FLAGS.task_id,
total_num_replicas=num_workers,
variable_averages=exp_moving_averager,
variables_to_average=variables_to_average)
batchnorm_updates = tf.get_collection(slim.ops.UPDATE_OPS_COLLECTION)
assert batchnorm_updates, 'Batchnorm updates are missing'
batchnorm_updates_op = tf.group(*batchnorm_updates)
# Add dependency to compute batchnorm_updates.
with tf.control_dependencies([batchnorm_updates_op]):
total_loss = tf.identity(total_loss)
# Compute gradients with respect to the loss.
grads = opt.compute_gradients(total_loss)
# Add histograms for gradients.
for grad, var in grads:
if grad is not None:
tf.summary.histogram(var.op.name + '/gradients', grad)
apply_gradients_op = opt.apply_gradients(grads, global_step=global_step)
with tf.control_dependencies([apply_gradients_op]):
train_op = tf.identity(total_loss, name='train_op')
# Get chief queue_runners, init_tokens and clean_up_op, which is used to
# synchronize replicas.
# More details can be found in sync_replicas_optimizer.
chief_queue_runners = [opt.get_chief_queue_runner()]
init_tokens_op = opt.get_init_tokens_op()
clean_up_op = opt.get_clean_up_op()
# Create a saver.
saver = tf.train.Saver()
# Build the summary operation based on the TF collection of Summaries.
summary_op = tf.merge_all_summaries()
# Build an initialization operation to run below.
init_op = tf.initialize_all_variables()
# We run the summaries in the same thread as the training operations by
# passing in None for summary_op to avoid a summary_thread being started.
# Running summaries and training operations in parallel could run out of
# GPU memory.
sv = tf.train.Supervisor(is_chief=is_chief,
logdir=FLAGS.train_dir,
init_op=init_op,
summary_op=None,
global_step=global_step,
saver=saver,
save_model_secs=FLAGS.save_interval_secs)
tf.logging.info('%s Supervisor' % datetime.now())
sess_config = tf.ConfigProto(
allow_soft_placement=True,
log_device_placement=FLAGS.log_device_placement)
# Get a session.
sess = sv.prepare_or_wait_for_session(target, config=sess_config)
# Start the queue runners.
queue_runners = tf.get_collection(tf.GraphKeys.QUEUE_RUNNERS)
sv.start_queue_runners(sess, queue_runners)
tf.logging.info('Started %d queues for processing input data.',
len(queue_runners))
if is_chief:
sv.start_queue_runners(sess, chief_queue_runners)
sess.run(init_tokens_op)
# Train, checking for Nans. Concurrently run the summary operation at a
# specified interval. Note that the summary_op and train_op never run
# simultaneously in order to prevent running out of GPU memory.
next_summary_time = time.time() + FLAGS.save_summaries_secs
while not sv.should_stop():
try:
start_time = time.time()
loss_value, step = sess.run([train_op, global_step])
assert not np.isnan(loss_value), 'Model diverged with loss = NaN'
if step > FLAGS.max_steps:
break
duration = time.time() - start_time
if step % 30 == 0:
examples_per_sec = FLAGS.batch_size / float(duration)
format_str = ('Worker %d: %s: step %d, loss = %.2f'
'(%.1f examples/sec; %.3f sec/batch)')
tf.logging.info(format_str %
(FLAGS.task_id, datetime.now(), step, loss_value,
examples_per_sec, duration))
# Determine if the summary_op should be run on the chief worker.
if is_chief and next_summary_time < time.time():
tf.logging.info('Running Summary operation on the chief.')
summary_str = sess.run(summary_op)
sv.summary_computed(sess, summary_str)
tf.logging.info('Finished running Summary operation.')
# Determine the next time for running the summary.
next_summary_time += FLAGS.save_summaries_secs
except:
if is_chief:
tf.logging.info('About to execute sync_clean_up_op!')
sess.run(clean_up_op)
raise
# Stop the supervisor. This also waits for service threads to finish.
sv.stop()
# Save after the training ends.
if is_chief:
saver.save(sess,
os.path.join(FLAGS.train_dir, 'model.ckpt'),
global_step=global_step)
def train(target, dataset, cluster_spec):
"""Train Inception on a dataset for a number of steps."""
# Number of workers and parameter servers are infered from the workers and ps
# hosts string.
num_workers = len(cluster_spec.as_dict()['worker'])
num_parameter_servers = len(cluster_spec.as_dict()['ps'])
# If no value is given, num_replicas_to_aggregate defaults to be the number of
# workers.
if FLAGS.num_replicas_to_aggregate == -1:
num_replicas_to_aggregate = num_workers
else:
num_replicas_to_aggregate = FLAGS.num_replicas_to_aggregate
# Both should be greater than 0 in a distributed training.
assert num_workers > 0 and num_parameter_servers > 0, (
' num_workers and '
'num_parameter_servers'
' must be > 0.')
# Choose worker 0 as the chief. Note that any worker could be the chief
# but there should be only one chief.
is_chief = (FLAGS.task_id == 0)
# Ops are assigned to worker by default.
with tf.device('/job:worker/task:%d' % FLAGS.task_id):
# Variables and its related init/assign ops are assigned to ps.
with slim.scopes.arg_scope(
[slim.variables.variable, slim.variables.global_step],
device=slim.variables.VariableDeviceChooser(
num_parameter_servers)):
# Create a variable to count the number of train() calls. This equals the
# number of updates applied to the variables.
global_step = slim.variables.global_step()
# Calculate the learning rate schedule.
num_batches_per_epoch = (dataset.num_examples_per_epoch() /
FLAGS.batch_size)
# Decay steps need to be divided by the number of replicas to aggregate.
decay_steps = int(num_batches_per_epoch *
FLAGS.num_epochs_per_decay /
num_replicas_to_aggregate)
# Decay the learning rate exponentially based on the number of steps.
lr = tf.train.exponential_decay(FLAGS.initial_learning_rate,
global_step,
decay_steps,
FLAGS.learning_rate_decay_factor,
staircase=True)
# Add a summary to track the learning rate.
tf.summary.scalar('learning_rate', lr)
# Create an optimizer that performs gradient descent.
opt = tf.train.RMSPropOptimizer(lr,
RMSPROP_DECAY,
momentum=RMSPROP_MOMENTUM,
epsilon=RMSPROP_EPSILON)
images, labels = image_processing.distorted_inputs(
dataset,
batch_size=FLAGS.batch_size,
num_preprocess_threads=FLAGS.num_preprocess_threads)
# Number of classes in the Dataset label set plus 1.
# Label 0 is reserved for an (unused) background class.
num_classes = dataset.num_classes() + 1
logits = inception.inference(images,
num_classes,
for_training=True)
# Add classification loss.
inception.loss(logits, labels)
# Gather all of the losses including regularization losses.
losses = tf.get_collection(slim.losses.LOSSES_COLLECTION)
losses += tf.get_collection(tf.GraphKeys.REGULARIZATION_LOSSES)
total_loss = tf.add_n(losses, name='total_loss')
if is_chief:
# Compute the moving average of all individual losses and the
# total loss.
loss_averages = tf.train.ExponentialMovingAverage(0.9,
name='avg')
loss_averages_op = loss_averages.apply(losses + [total_loss])
# Attach a scalar summmary to all individual losses and the total loss;
# do the same for the averaged version of the losses.
for l in losses + [total_loss]:
loss_name = l.op.name
# Name each loss as '(raw)' and name the moving average version of the
# loss as the original loss name.
tf.summary.scalar(loss_name + ' (raw)', l)
tf.summary.scalar(loss_name, loss_averages.average(l))
# Add dependency to compute loss_averages.
with tf.control_dependencies([loss_averages_op]):
total_loss = tf.identity(total_loss)
# Track the moving averages of all trainable variables.
# Note that we maintain a 'double-average' of the BatchNormalization
# global statistics.
# This is not needed when the number of replicas are small but important
# for synchronous distributed training with tens of workers/replicas.
exp_moving_averager = tf.train.ExponentialMovingAverage(
inception.MOVING_AVERAGE_DECAY, global_step)
variables_to_average = (tf.trainable_variables() +
tf.moving_average_variables())
# Add histograms for model variables.
for var in variables_to_average:
tf.summary.histogram(var.op.name, var)
# Create synchronous replica optimizer.
opt = tf.train.SyncReplicasOptimizer(
opt,
replicas_to_aggregate=num_replicas_to_aggregate,
replica_id=FLAGS.task_id,
total_num_replicas=num_workers,
variable_averages=exp_moving_averager,
variables_to_average=variables_to_average)
batchnorm_updates = tf.get_collection(
slim.ops.UPDATE_OPS_COLLECTION)
assert batchnorm_updates, 'Batchnorm updates are missing'
batchnorm_updates_op = tf.group(*batchnorm_updates)
# Add dependency to compute batchnorm_updates.
with tf.control_dependencies([batchnorm_updates_op]):
total_loss = tf.identity(total_loss)
# Compute gradients with respect to the loss.
grads = opt.compute_gradients(total_loss)
# Add histograms for gradients.
for grad, var in grads:
if grad is not None:
tf.summary.histogram(var.op.name + '/gradients', grad)
apply_gradients_op = opt.apply_gradients(grads,
global_step=global_step)
with tf.control_dependencies([apply_gradients_op]):
train_op = tf.identity(total_loss, name='train_op')
# Get chief queue_runners, init_tokens and clean_up_op, which is used to
# synchronize replicas.
# More details can be found in sync_replicas_optimizer.
chief_queue_runners = [opt.get_chief_queue_runner()]
init_tokens_op = opt.get_init_tokens_op()
clean_up_op = opt.get_clean_up_op()
# Create a saver.
saver = tf.train.Saver()
# Build the summary operation based on the TF collection of Summaries.
summary_op = tf.merge_all_summaries()
# Build an initialization operation to run below.
init_op = tf.initialize_all_variables()
# We run the summaries in the same thread as the training operations by
# passing in None for summary_op to avoid a summary_thread being started.
# Running summaries and training operations in parallel could run out of
# GPU memory.
sv = tf.train.Supervisor(is_chief=is_chief,
logdir=FLAGS.train_dir,
init_op=init_op,
summary_op=None,
global_step=global_step,
saver=saver,
save_model_secs=FLAGS.save_interval_secs)
tf.logging.info('%s Supervisor' % datetime.now())
sess_config = tf.ConfigProto(
allow_soft_placement=True,
log_device_placement=FLAGS.log_device_placement)
# Get a session.
sess = sv.prepare_or_wait_for_session(target, config=sess_config)
# Start the queue runners.
queue_runners = tf.get_collection(tf.GraphKeys.QUEUE_RUNNERS)
sv.start_queue_runners(sess, queue_runners)
tf.logging.info('Started %d queues for processing input data.',
len(queue_runners))
if is_chief:
sv.start_queue_runners(sess, chief_queue_runners)
sess.run(init_tokens_op)
# Train, checking for Nans. Concurrently run the summary operation at a
# specified interval. Note that the summary_op and train_op never run
# simultaneously in order to prevent running out of GPU memory.
next_summary_time = time.time() + FLAGS.save_summaries_secs
while not sv.should_stop():
try:
start_time = time.time()
loss_value, step = sess.run([train_op, global_step])
assert not np.isnan(
loss_value), 'Model diverged with loss = NaN'
if step > FLAGS.max_steps:
break
duration = time.time() - start_time
if step % 30 == 0:
examples_per_sec = FLAGS.batch_size / float(duration)
format_str = ('Worker %d: %s: step %d, loss = %.2f'
'(%.1f examples/sec; %.3f sec/batch)')
tf.logging.info(
format_str %
(FLAGS.task_id, datetime.now(), step, loss_value,
examples_per_sec, duration))
# Determine if the summary_op should be run on the chief worker.
if is_chief and next_summary_time < time.time():
tf.logging.info(
'Running Summary operation on the chief.')
summary_str = sess.run(summary_op)
sv.summary_computed(sess, summary_str)
tf.logging.info('Finished running Summary operation.')
# Determine the next time for running the summary.
next_summary_time += FLAGS.save_summaries_secs
except:
if is_chief:
tf.logging.info('About to execute sync_clean_up_op!')
sess.run(clean_up_op)
raise
# Stop the supervisor. This also waits for service threads to finish.
sv.stop()
# Save after the training ends.
if is_chief:
saver.save(sess,
os.path.join(FLAGS.train_dir, 'model.ckpt'),
global_step=global_step)