What's New in TensorFlow 2.0

Saving and loading models

After training, it can be extremely useful to save the model for later use. Decoupling the training and inference pipelines is a good idea in many use cases. From a developer's perspective, a model can be abstracted out as a black box that accepts a set of inputs and returns some outputs. Saving a model, then, is nothing but exporting an artifact that represents this black box. Restoring or loading models then becomes the process of using this black box to perform some real work. This can also be understood as the process of serializing and deserializing the model black box.

TF 2.0 supports saving and restoring a model in multiple modes:

• Model architecture only (Keras)
• Model weights only (Keras)
• The entire model: ...