def inference_on_image(img_path):
with tf.Graph().as_default():
num_classes = 1001
coder = build_image_data.ImageCoder()
image_buffer, _, _ = build_image_data._process_image(img_path, coder)
image = image_processing.image_preprocessing(
image_buffer, 0, False) # image -> (299, 299, 3)
image = tf.expand_dims(image, 0) # (299, 299,3) -> (1, 299, 299, 3)
logits, _ = inception.inference(image,
num_classes,
for_training=False,
restore_logits=True)
saver = tf.train.Saver()
with tf.Session() as tf_session:
ckpt = tf.train.get_checkpoint_state(checkpoint_dir)
if ckpt and ckpt.model_checkpoint_path:
if os.path.isabs(ckpt.model_checkpoint_path):
# Restores from checkpoint with absolute path.
saver.restore(tf_session, ckpt.model_checkpoint_path)
# saver.restore(tf_session, "/tmp/train_back/model.ckpt-640")
else:
# Restores from checkpoint with relative path.
saver.restore(
tf_session,
os.path.join(FLAGS.checkpoint_dir,
ckpt.model_checkpoint_path))
l = tf_session.run([logits])
return l
def inference_on_multi_image():
print("total image size {} ".format(file_size))
total_batch_size = int(file_size / batch_size + 1)
logit_list = []
for n in range(total_batch_size):
print("step :{} / {}".format(n + 1, total_batch_size))
mini_batch = img_file_list[n * batch_size:(n + 1) * batch_size]
mini_adarr = np.ndarray(shape=(0, 299, 299, 3))
with tf.Graph().as_default():
num_classes = 1001
coder = build_image_data.ImageCoder()
for i, image in enumerate(mini_batch):
image_buffer, _, _ = build_image_data._process_image(
my_image_path + image, coder)
image = image_processing.image_preprocessing(
image_buffer, 0, False) # image -> (299, 299, 3)
image = tf.expand_dims(image,
0) # (299, 299,3) -> (1, 299, 299, 3)
mini_adarr = tf.concat([mini_adarr, image], 0)
logits, _ = inception.inference(mini_adarr,
num_classes,
for_training=False,
restore_logits=True)
saver = tf.train.Saver()
with tf.Session() as tf_session:
ckpt = tf.train.get_checkpoint_state(checkpoint_dir)
if ckpt and ckpt.model_checkpoint_path:
if os.path.isabs(ckpt.model_checkpoint_path):
# Restores from checkpoint with absolute path.
saver.restore(tf_session, ckpt.model_checkpoint_path)
# saver.restore(tf_session, "/tmp/train_back/model.ckpt-640")
else:
# Restores from checkpoint with relative path.
saver.restore(
tf_session,
os.path.join(FLAGS.checkpoint_dir,
ckpt.model_checkpoint_path))
l = tf_session.run([logits])
for ll in l[0]:
logit_list.append(ll)
return logit_list
def train(target, dataset, cluster_spec, ctx):
"""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)
if FLAGS.input_mode == 'spark':
def feed_dict(mgr, batch_size):
tmp = TFNode.next_batch(mgr, batch_size)
# extract TFRecords, since tmp array is [(TFRecord, None)]
tfrecords = []
for elem in tmp:
tfrecords.append(str(elem[0]))
return tfrecords
batch = tf.placeholder(
tf.string,
[FLAGS.batch_size / FLAGS.num_preprocess_threads])
# The following is adapted from image_processing.py to remove Readers/QueueRunners.
# Note: this removes the RandomShuffledQueue, so the incoming data is not shuffled.
# Presumably, this could be done on the Spark side or done in additional TF code.
examples = tf.unpack(batch)
images, labels = [], []
for example_serialized in examples:
for thread_id in range(FLAGS.num_preprocess_threads):
# Parse a serialized Example proto to extract the image and metadata.
image_buffer, label_index, bbox, _ = image_processing.parse_example_proto(
example_serialized)
image = image_processing.image_preprocessing(
image_buffer, bbox, train, thread_id)
images.append(image)
labels.append(label_index)
height = FLAGS.image_size
width = FLAGS.image_size
depth = 3
images = tf.cast(images, tf.float32)
images = tf.reshape(
images, shape=[FLAGS.batch_size, height, width, depth])
tf.summary.image('images', images)
labels = tf.reshape(labels, [FLAGS.batch_size])
else:
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,
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()
# Create a saver.
saver = tf.train.Saver()
# Build the summary operation based on the TF collection of Summaries.
summary_op = tf.summary.merge_all()
# Build an initialization operation to run below.
init_op = tf.global_variables_initializer()
# 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.
summary_writer = tf.summary.FileWriter(
"tensorboard_%d" % (ctx.worker_num),
graph=tf.get_default_graph())
sv = tf.train.Supervisor(is_chief=is_chief,
logdir=FLAGS.train_dir,
init_op=init_op,
summary_op=None,
global_step=global_step,
summary_writer=summary_writer,
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()
if FLAGS.input_mode == 'spark':
tmp = feed_dict(
ctx.mgr,
FLAGS.batch_size / FLAGS.num_preprocess_threads)
feed = {batch: tmp}
loss_value, step = sess.run([train_op, global_step],
feed_dict=feed)
else:
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 FLAGS.input_mode == 'tf' and 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!')
raise
# Stop the TFNode data feed
if FLAGS.input_mode == 'spark':
TFNode.terminate(ctx.mgr)
# 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, ctx):
"""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)
if FLAGS.input_mode == 'spark':
def feed_dict(mgr, batch_size):
tmp = TFNode.next_batch(mgr, batch_size)
# extract TFRecords, since tmp array is [(TFRecord, None)]
tfrecords = []
for elem in tmp:
tfrecords.append(str(elem[0]))
return tfrecords
batch = tf.placeholder(tf.string, [FLAGS.batch_size/FLAGS.num_preprocess_threads])
# The following is adapted from image_processing.py to remove Readers/QueueRunners.
# Note: this removes the RandomShuffledQueue, so the incoming data is not shuffled.
# Presumably, this could be done on the Spark side or done in additional TF code.
examples = tf.unpack(batch)
images, labels = [], []
for example_serialized in examples:
for thread_id in range(FLAGS.num_preprocess_threads):
# Parse a serialized Example proto to extract the image and metadata.
image_buffer, label_index, bbox, _ = image_processing.parse_example_proto(example_serialized)
image = image_processing.image_preprocessing(image_buffer, bbox, train, thread_id)
images.append(image)
labels.append(label_index)
height = FLAGS.image_size
width = FLAGS.image_size
depth = 3
images = tf.cast(images, tf.float32)
images = tf.reshape(images, shape=[FLAGS.batch_size, height, width, depth])
tf.summary.image('images', images)
labels = tf.reshape(labels, [FLAGS.batch_size])
else:
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,
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()
# Create a saver.
saver = tf.train.Saver()
# Build the summary operation based on the TF collection of Summaries.
summary_op = tf.summary.merge_all()
# Build an initialization operation to run below.
init_op = tf.global_variables_initializer()
# 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.
summary_writer = tf.summary.FileWriter("tensorboard_%d" %(ctx.worker_num), graph=tf.get_default_graph())
sv = tf.train.Supervisor(is_chief=is_chief,
logdir=FLAGS.train_dir,
init_op=init_op,
summary_op=None,
global_step=global_step,
summary_writer=summary_writer,
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()
if FLAGS.input_mode == 'spark':
tmp = feed_dict(ctx.mgr, FLAGS.batch_size/FLAGS.num_preprocess_threads)
feed = {batch: tmp}
loss_value, step = sess.run([train_op, global_step], feed_dict=feed)
else:
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 FLAGS.input_mode == 'tf' and 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!')
raise
# Stop the TFNode data feed
if FLAGS.input_mode == 'spark':
TFNode.terminate(ctx.mgr)
# 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)