Tensorflow is making Keras and eager execution the center of its new API in version 2: https://medium.com/tensorflow/standardizing-on-keras-guidance-on-high-level-apis-in-tensorflow-2-0-bad2b04c819a While estimators are going to be supported, they do not support eager execution (They always run in graph mode).

Per this guide, it's best to run code that runs both in eager mode and graph mode. I think we can extend our estimator classes to support their execution in eager mode, i.e., we can have one eager execution training script that runs just like estimator.train but in eager mode. This allows for easier debugging of our programs and lets us to easily switch the same model training/evaluation/prediction to graph mode.

any feedback is welcome