As we explained in Chapter 3, TensorFlow Graph Architecture, representing computations using DataFlow graphs has several advantages in terms of model portability since a graph is a language-agnostic representation of the computation.
SavedModel is a universal serialization format for TensorFlow models that extends the TensorFlow standard graph representation by creating a language-agnostic representation for the computation that is recoverable and hermetic. This representation has been designed not only to carry the graph description and values (like the standard graph) but also to offer additional features that were designed to simplify the usage of the trained models in heterogeneous production environments.
TensorFlow 2.0 has been designed with simplicity in mind. This design choice is visible in the following diagram, where it is possible...
