https://towardsdatascience.com/how-to-quickly-build-a-tensorflow-training-pipeline-15e9ae4d78a0

Tensorflow is great. Really, you can do everything imaginable. You can turn zebras into horses with it. However, Tensorflow’s code examples generally tend to gloss over to get data into your model: they either sometimes naively assume that someone else did the hard work for you and serialized the data into Tensorflow’s native format, or showcase unreasonably slow methods that would have a GPU idling away with shockingly low performance. Also oftentimes the code is very hacky and difficult to follow. So I thought it might be useful to show a small, self contained example that handles both training and efficient data pipelining on a nontrivial example.

In typical Tensorflow fashion, there are many ways that you could get your data pipeline set up. This guide will quickly list the top 3, and show you how to use a compromise that gets you that go-to solution that is very easy to code and blazingly fast for the for 80% of the use cases.