关于Tensorflow中的入队顺序

Question

I am relatively a beginner of Python and Tensorflow. And when I was learning the threading and queues part of Tensorflow, I was a little confused.

So when we use multi-threads (generated by QueueRunner) to enqueue one queue from the data source, what the enqueue order will be? Will the data in the queue keep the original order in the data source? If so, how to achieve this with multi-threads? If not, why do we need RandomShuffleQueue (if the enqueue order is not fixed, shuffle the dequeue order seems a little redundant)?

Thank you in advance.

Answer 1

First of all, you should not use queues/queue runners for input pipeline anymore. Use Datasets API instead. You should find it more intuitive and much simpler.

To answer your questions:

Re: So when we use multi-threads (generated by QueueRunner) to enqueue one queue from the data source, what the enqueue order will be?

The order is not defined. It depends on thread scheduling.

Re: Will the data in the queue keep the original order in the data source?

The data source is usually a set of files and usually a given file is processed by a single thread and a single reader. If that is the case, order of examples coming from one file will be preserved (unless you purposefully shuffle them, eg using RandomShuffleQueue instead of FIFOQueue)

Re: why do we need RandomShuffleQueue?

In the common scenario above, if you don't shuffle examples, their order will be preserved. You might want to shuffle your examples for training because they might follow some order within a file. Also, while the order across files is not deterministic, it is far from uniform random. You might want to achieve (a good approximation to) uniform random order across your entire data set.