How does tf.train.batch create a batch

Question

In the CIFAR-10 TensorFlow tutorial I encountered the following line:

images, label_batch = tf.train.batch(
        [image, label],
        batch_size=batch_size,
        num_threads=num_preprocess_threads,
        capacity=min_queue_examples + 3 * batch_size)

The function tf.train.batch() seems to be taking as an input only one image and one label. How does it then create a batch with multiple images?

Answer 1

It takes in input the pair [image, label] that, yes, it's a single pair. tf.train.batch, however, creates a queue internally. The num_threads threads will accumulate the pairs into the queue until capacity is reached.

The images, label_batch are, in fact, dequeue operations.

Remember that you're defining a computational graph, therefore the pair [image, label] represent two nodes of the graph and the the various real pair of image, label of your training set, will flow through these nodes. In this way, the tf.train.batch can capture the flow of the images and labels and fill the queue.