TensorFlow shuffle_batch not working

Question

import tensorflow as tf
sess = tf.Session()

def add_to_batch(image):

    print('Adding to batch')
    image_batch = tf.train.shuffle_batch([image],batch_size=5,capacity=11,min_after_dequeue=1,num_threads=1)

    # Add to summary
    tf.image_summary('images',image_batch)

    return image_batch

def get_batch():

    # Create filename queue of images to read
    filenames = [('/media/jessica/Jessica/TensorFlow/Practice/unlabeled_data_%d.png' % i) for i in range(11)]
    filename_queue = tf.train.string_input_producer(filenames)
    reader = tf.WholeFileReader()
    key, value = reader.read(filename_queue)

    # Read and process image
    my_image = tf.image.decode_png(value)
    my_image_float = tf.cast(my_image,tf.float32)
    image_mean = tf.reduce_mean(my_image_float)
    my_noise = tf.random_normal([96,96,3],mean=image_mean)
    my_image_noisy = my_image_float + my_noise
    print('Reading images')

    return add_to_batch(my_image_noisy)

def main ():

    sess.run(tf.initialize_all_variables())
    tf.train.start_queue_runners(sess=sess)
    writer = tf.train.SummaryWriter('/media/jessica/Jessica/TensorFlow/Practice/summary_logs', graph_def=sess.graph_def)
    merged = tf.merge_all_summaries()
    images = get_batch()
    summary_str = sess.run(merged)
    writer.add_summary(summary_str)

Hi,

I am trying to build a simple neural network in TensorFlow. I am trying to load my input images in batches. Now I am testing the code with 11 images and batch_size = 5. Eventually I will be working with 100000 images.

This bit of code is modified from TensorFlow's cifar10.py example. For some reason my code stops (doesn't terminate, it just hangs there) at tf.train.shuffle_batch([image],batch_size=5,capacity=1,min_after_dequeue=1,num_threads=1)

I've tried different combinations of batch_size, capacity, min_after_dequeue etc. but I still can't figure out what is going wrong.

Any help will be much appreciated! Thank you!

Answer 1

It looks like the problem arises because the statement

tf.train.start_queue_runners(sess=sess)

...executes before any queue runners have been created. If you move this line after images = get_batch(), your program should work.

What is the problem here? The tf.train.shuffle_batch() function internally uses a tf.RandomShuffleQueue to produce a randomized batch. Currently, the only way to get elements into that queue is to run a step that invokes the q.enqueue() op. To make this easier, TensorFlow has a notion of "queue runners" that are implicitly collected as you build your graph, and then started with a call to tf.train.start_queue_runners(). However, calling tf.train.start_queue_runners() only starts the queue runners that have been defined at that point in time, so it must come after the code that creates the queue runners.