Input raw data (e.g., text, image address). There are many different implementations of queue.
feature-label pair: [enqueue-op] [fetch-func] [slice-input-producer] [string-input-producer]
image-image pair: [slice-input-producer][string-input-producer]
A more elegant way is converting raw data to tfrecord format.
Tips:
setting large number_of_threading (e.g., 10) is helpful.
shuffle the training samples to avoid homogenuity when necessary.
place the training data in local disk instead of removable disk (consider I/O speed).