The best way to start is by having a look at the programming environment:

We are obviously interested in the Python stack, and we will mainly focus on layers and metrics. Datasets are interesting, but lots of them are from external contributions, and some are targeted for removal. The scikit learn API is considered as more future-proof, so we won't look at it.

Estimators are the higher-level APIs, but they are not as well developed as the one from scikit learn. As we develop new networks, being able to debug them and check what they have inside their gut is easier in the middle API than the top one, although the fact that all tensors have names makes it possible to get this information outside of the Estimator API.

Lots of online tutorials are still directly using the lower API, and we used it in our regression example by calling tf.matmult directly. We think it is better to use the middle- or the high-level API than the others, even if they may sometimes seem more flexible and closer to what you think you need.