Tensorflow 2.0: Keras is not (yet) a simplified interface to Tensorflow

In Tensorflow 2.0 Keras will be the default high-level API for building and training machine learning models, hence complete compatibility between a model defined using the old tf.layers and the new tf.Keras.layers is expected. In version 2 of the popular machine learning framework the eager execution will be enabled by default although the static graph definition session execution will be still supported (but hidden a little bit). In this post, you'll see that the compatibility between a model defined using tf.layers and tf.keras.layers is not always guaranteed when using the graph definition session execution, but it works as expected if the eager execution is enabled (at least from my tests). The post is organized as follows: definition of the common data input pipeline, definition of the same model using both tf.layers and tf.keras.layers, The model we're going to use to highlight the differences between the 2 versions is a simple binary classifier. Instead of using a single tf.data.Dataset object with both the positive and negative classes inside, we want to use the variable sharing feature of Tensorflow (and thus of Keras models) to feed first the positive and then the negative ones in order to test also if the behavior of the variable sharing still works as we're used to.