train_network.py

# Initialize placeholders for feeding in to the queue
self.queue_inputs       = tf.placeholder(tf.float32, shape=[None, self.config.seq_length, self.config.image_size, self.config.image_size], name="queue_inputs")
self.queue_targets      = tf.placeholder(tf.uint8, shape=[None, self.config.seq_length], name="queue_targets")

min_after_dequeue = 10000
capacity = min_after_dequeue + 3 * self.config.batch_size

q = tf.FIFOQueue(
    capacity=15000,
    dtypes=[tf.float32, tf.uint8],
    shapes=[[self.config.seq_length, self.config.image_size, self.config.image_size], [self.config.seq_length]]
)

self.enqueue_op = q.enqueue_many([self.queue_inputs, self.queue_targets])
self.examples_in_queue = q.size()

# This must be the input for the training operation
self.inputs_batch_queue, self.targets_batch_queue = q.dequeue_many(self.config.batch_size)

# Placeholders for training and evaluation
self.inputs_batch  = tf.placeholder_with_default(self.inputs_batch_queue, [None, self.config.seq_length, self.config.image_size, self.config.image_size], name="model_inputs")
self.targets_batch = tf.placeholder_with_default(self.targets_batch_queue, [None, self.config.seq_length], name="model_targets")

# ...

############################################################################
############################################################################
############################################################################

# Coordinator for threads
coord = tf.train.Coordinator()

# Start a thread to enqueue data asynchronously, and hide I/O latency.
t = threading.Thread(target=load_and_enqueue, args=(
    config.input_dir, sess, coord,
    model.enqueue_op, model.queue_inputs,
    model.queue_targets, config.num_train, 100, config.rewrite_targets,
))
t.setDaemon(True)
t.start()