TensorFlow Saving and Loading Models 3

Question 1

What does setting include_optimizer=False in model.save() affect?

Accepted Answer

Excludes the optimizer state, making the file smaller but unable to resume training exactly

Answer

With include_optimizer=False the saved file omits the optimizer's moment variables and iteration count, reducing file size but requiring recompilation before resuming training.

Question 2

What type of object does tf.saved_model.load() return?

Accepted Answer

A trackable object containing the model's saved functions and variables

Answer

tf.saved_model.load() returns a low-level trackable object (AutoTrackable), not a Keras Model, so Keras-specific methods like .fit() are unavailable.

Question 3

Why should serving methods in a SavedModel be decorated with @tf.function?

Accepted Answer

It traces the Python function into a TensorFlow graph that can be serialized and run without Python

Answer

@tf.function triggers graph tracing, creating a ConcreteFunction that is serialized inside the SavedModel and can execute in any TensorFlow runtime without a Python interpreter.

Question 4

Which serialization format is required by TensorFlow Serving for model deployment?

Accepted Answer

SavedModel directory format

Answer

TensorFlow Serving is built to load and serve models from the SavedModel directory format, which exposes named signatures for gRPC and REST inference endpoints.

Question 5

What is the purpose of the signatures attribute on a loaded SavedModel object?

Accepted Answer

Providing named callable entry points used by serving infrastructure

Answer

The signatures dict maps string keys (e.g., 'serving_default') to ConcreteFunction objects that serving systems call to run predictions.

TensorFlow Practice Test

TensorFlow Practice Test

TensorFlow Saving and Loading Models 3