We consistently see issues when users use `tf.keras.layers`, which are fixed when they switch to the `tf.layers` equivalent. Related to #1.
We consistently see issues when users use
tf.keras.layers, which are fixed when they switch to thetf.layersequivalent.Related to #1.