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 train(create_tensor_dict_fn,
create_model_fn,
train_config,
master,
task,
num_clones,
worker_replicas,
clone_on_cpu,
ps_tasks,
worker_job_name,
is_chief,
train_dir,
graph_hook_fn=None):
"""Training function for detection models.
Args:
create_tensor_dict_fn: a function to create a tensor input dictionary.
create_model_fn: a function that creates a DetectionModel and generates
losses.
train_config: a train_pb2.TrainConfig protobuf.
master: BNS name of the TensorFlow master to use.
task: The task id of this training instance.
num_clones: The number of clones to run per machine.
worker_replicas: The number of work replicas to train with.
clone_on_cpu: True if clones should be forced to run on CPU.
ps_tasks: Number of parameter server tasks.
worker_job_name: Name of the worker job.
is_chief: Whether this replica is the chief replica.
train_dir: Directory to write checkpoints and training summaries to.
graph_hook_fn: Optional function that is called after the inference graph is
built (before optimization). This is helpful to perform additional changes
to the training graph such as adding FakeQuant ops. The function should
modify the default graph.
Raises:
ValueError: If both num_clones > 1 and train_config.sync_replicas is true.
"""
detection_model = create_model_fn()
data_augmentation_options = [
preprocessor_builder.build(step)
for step in train_config.data_augmentation_options
]
with tf.Graph().as_default():
# Build a configuration specifying multi-GPU and multi-replicas.
deploy_config = model_deploy.DeploymentConfig(
num_clones=num_clones,
clone_on_cpu=clone_on_cpu,
replica_id=task,
num_replicas=worker_replicas,
num_ps_tasks=ps_tasks,
worker_job_name=worker_job_name)
# Place the global step on the device storing the variables.
with tf.device(deploy_config.variables_device()):
global_step = slim.create_global_step()
if num_clones != 1 and train_config.sync_replicas:
raise ValueError('In Synchronous SGD mode num_clones must ',
'be 1. Found num_clones: {}'.format(num_clones))
batch_size = train_config.batch_size // num_clones
if train_config.sync_replicas:
batch_size //= train_config.replicas_to_aggregate
with tf.device(deploy_config.inputs_device()):
input_queue = create_input_queue(
batch_size, create_tensor_dict_fn,
train_config.batch_queue_capacity,
train_config.num_batch_queue_threads,
train_config.prefetch_queue_capacity,
data_augmentation_options)
# Gather initial summaries.
# TODO(rathodv): See if summaries can be added/extracted from global tf
# collections so that they don't have to be passed around.
summaries = set(tf.get_collection(tf.GraphKeys.SUMMARIES))
global_summaries = set([])
model_fn = functools.partial(_create_losses,
create_model_fn=create_model_fn,
train_config=train_config)
clones = model_deploy.create_clones(deploy_config, model_fn,
[input_queue])
first_clone_scope = clones[0].scope
if graph_hook_fn:
with tf.device(deploy_config.variables_device()):
graph_hook_fn()
# 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)
with tf.device(deploy_config.optimizer_device()):
training_optimizer, optimizer_summary_vars = optimizer_builder.build(
train_config.optimizer)
for var in optimizer_summary_vars:
tf.summary.scalar(var.op.name, var, family='LearningRate')
sync_optimizer = None
if train_config.sync_replicas:
training_optimizer = tf.train.SyncReplicasOptimizer(
training_optimizer,
replicas_to_aggregate=train_config.replicas_to_aggregate,
total_num_replicas=worker_replicas)
sync_optimizer = training_optimizer
with tf.device(deploy_config.optimizer_device()):
regularization_losses = (
None if train_config.add_regularization_loss else [])
total_loss, grads_and_vars = model_deploy.optimize_clones(
clones,
training_optimizer,
regularization_losses=regularization_losses)
total_loss = tf.check_numerics(total_loss,
'LossTensor is inf or nan.')
# Optionally multiply bias gradients by train_config.bias_grad_multiplier.
if train_config.bias_grad_multiplier:
biases_regex_list = ['.*/biases']
grads_and_vars = variables_helper.multiply_gradients_matching_regex(
grads_and_vars,
biases_regex_list,
multiplier=train_config.bias_grad_multiplier)
# Optionally freeze some layers by setting their gradients to be zero.
if train_config.freeze_variables:
grads_and_vars = variables_helper.freeze_gradients_matching_regex(
grads_and_vars, train_config.freeze_variables)
# Optionally clip gradients
if train_config.gradient_clipping_by_norm > 0:
with tf.name_scope('clip_grads'):
grads_and_vars = slim.learning.clip_gradient_norms(
grads_and_vars, train_config.gradient_clipping_by_norm)
# Create gradient updates.
grad_updates = training_optimizer.apply_gradients(
grads_and_vars, global_step=global_step)
update_ops.append(grad_updates)
update_op = tf.group(*update_ops, name='update_barrier')
with tf.control_dependencies([update_op]):
train_tensor = tf.identity(total_loss, name='train_op')
# Add summaries.
for model_var in slim.get_model_variables():
global_summaries.add(
tf.summary.histogram('ModelVars/' + model_var.op.name,
model_var))
for loss_tensor in tf.losses.get_losses():
global_summaries.add(
tf.summary.scalar('Losses/' + loss_tensor.op.name,
loss_tensor))
global_summaries.add(
tf.summary.scalar('Losses/TotalLoss', tf.losses.get_total_loss()))
# 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))
summaries |= global_summaries
# Merge all summaries together.
summary_op = tf.summary.merge(list(summaries), name='summary_op')
# Soft placement allows placing on CPU ops without GPU implementation.
session_config = tf.ConfigProto(allow_soft_placement=True,
log_device_placement=False)
# Save checkpoints regularly.
keep_checkpoint_every_n_hours = train_config.keep_checkpoint_every_n_hours
saver = tf.train.Saver(
keep_checkpoint_every_n_hours=keep_checkpoint_every_n_hours)
# Create ops required to initialize the model from a given checkpoint.
init_fn = None
if train_config.fine_tune_checkpoint:
if not train_config.fine_tune_checkpoint_type:
# train_config.from_detection_checkpoint field is deprecated. For
# backward compatibility, fine_tune_checkpoint_type is set based on
# from_detection_checkpoint.
if train_config.from_detection_checkpoint:
train_config.fine_tune_checkpoint_type = 'detection'
else:
train_config.fine_tune_checkpoint_type = 'classification'
var_map = detection_model.restore_map(
fine_tune_checkpoint_type=train_config.
fine_tune_checkpoint_type,
load_all_detection_checkpoint_vars=(
train_config.load_all_detection_checkpoint_vars))
available_var_map = (
variables_helper.get_variables_available_in_checkpoint(
var_map,
train_config.fine_tune_checkpoint,
include_global_step=False))
init_saver = tf.train.Saver(available_var_map)
def initializer_fn(sess):
init_saver.restore(sess, train_config.fine_tune_checkpoint)
init_fn = initializer_fn
slim.learning.train(
train_tensor,
logdir=train_dir,
master=master,
is_chief=is_chief,
session_config=session_config,
startup_delay_steps=train_config.startup_delay_steps,
init_fn=init_fn,
summary_op=summary_op,
number_of_steps=(train_config.num_steps
if train_config.num_steps else None),
save_summaries_secs=120,
sync_optimizer=sync_optimizer,
saver=saver)
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
if FLAGS.quantize_delay >= 0:
tf.contrib.quantize.create_training_graph(
quant_delay=FLAGS.quantize_delay)
#########################################
# 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)