Read big train/validation/test datasets in tensorflow

Question

what's the right way to load more than one big datasets into tensor flow?

I have three big datasets(files), for train, validate and test respectively. I can successfully load training set thru tf.train.string_input_producer, and feeds it into a tf.train.shuffle_batch object. Then I can iteratively get batch of data to optimize my model.

But, I got stuck when trying to load my validation set by the same way, the program keeps saying "OutOfRange Error" even I didn't set num_epochs in string_input_producer.

Can anyone shed some lights on it? And besides this, I am also thinking what's the right approach to do training/validation in tensorflow? Actually, I didn't see any examples (I searched a lot) which have do both train and test on a big data set. It's so strange to me ...

Code snippet below.

def extract_validationset(filename, batch_size):
  with tf.device("/cpu:0"):
    queue = tf.train.string_input_producer([filename])
    reader = tf.TextLineReader()
    _, line = reader.read(queue)

    line = tf.decode_csv(...)
    label = line[0]
    feature = tf.pack(list(line[1:]))

    l, f = tf.train.batch([label, feature], batch_size=batch_size, num_threads=8)
    return l, f

def extract_trainset(train, batch_size):
  with tf.device("/cpu:0"):
    train_files = tf.train.string_input_producer([train])
    reader = tf.TextLineReader()
    _, train_line = reader.read(train_files)

    train_line = tf.decode_csv(...)

    l, f = tf.train.shuffle_batch(...,
  batch_size=batch_size, capacity=50000, min_after_dequeue=10000,  num_threads=8)
  return l, f

....

label_batch, feature_batch = extract_trainset("train", batch_size)
label_eval, feature_eval = extract_validationset("test", batch_size)

with tf.Session() as sess:
  tf.initialize_all_variables().run()
  coord = tf.train.Coordinator()
  threads = tf.train.start_queue_runners(coord=coord)

  # Loop through training steps.
  for step in xrange(int(num_epochs * train_size) // batch_size):
    feature, label = sess.run([feature_batch, label_batch])
    feed_dict = {train_data_node: feature, train_labels_node: label}

    _, l, predictions = sess.run([optimizer, loss, evaluation], feed_dict=feed_dict)

    # after EVAL_FREQUENCY steps, do evaluation on whole test set
    if step % EVAL_FREQUENCY == 0:
      for step in xrange(steps_per_epoch):
      f, l = sess.run([feature_eval, label_eval])
      true_count += sess.run(evaluation, feed_dict={train_data_node: f, train_labels_node: l})

    print('Precision @ 1: %0.04f' % true_count / num_examples)

<!---- ERROR ---->
tensorflow.python.framework.errors.OutOfRangeError: FIFOQueue '_5_batch/fifo_queue' is closed and has insufficient elements (requested 334, current size 0)
 [[Node: batch = QueueDequeueMany[component_types=[DT_FLOAT, DT_FLOAT], timeout_ms=-1, _device="/job:localhost/replica:0/task:0/cpu:0"](batch/fifo_queue, batch/n)]]

Caused by op u'batch', defined at:

Answer 1

This is maybe late but I had the same problem. In my case I was foolishly calling sess.run after I had closed shop with coord.request_stop(), coord.join_threads().

Maybe you have something like coord.request_stop() that is run in your "train" code, closing the queues for when you try to load your validation data.