SavedModel format in TensorFlow - Practice Problems & Coding Challenges

Challenge - 5 Problems

🧠 Conceptual

intermediate

Understanding the SavedModel directory structure

Which of the following best describes the main components inside a TensorFlow SavedModel directory?

A A Python script that recreates the model architecture and loads weights from a checkpoint.

B A single '.h5' file that stores the entire model including architecture and weights.

C A JSON file describing the model architecture and a separate binary file for weights.

D A directory containing a 'saved_model.pb' file and a 'variables' folder with model weights.

💡 Hint

Think about what files TensorFlow saves when you use model.save('path') with the default format.

The SavedModel format saves the model as a directory containing a saved_model.pb file which holds the model architecture and computation graph, plus a variables folder that contains the trained weights.

❓ Predict Output

intermediate

Output of loading a SavedModel

What will be the output of the following code snippet?

import tensorflow as tf
model = tf.keras.Sequential([tf.keras.layers.Dense(1)])
model.save('my_model')
loaded_model = tf.keras.models.load_model('my_model')
print(type(loaded_model))

A <class 'tensorflow.keras.models.Sequential'>

B <class 'tensorflow.saved_model.LoadedModel'>

C AttributeError: module 'tensorflow.keras.models' has no attribute 'load_model'

D <class 'tensorflow.keras.models.Model'>

💡 Hint

Consider what type of object load_model returns when loading a Keras model saved in SavedModel format.

The load_model function returns a Keras Sequential object, which is of type tensorflow.keras.models.Sequential.

❓ Model Choice

advanced

Choosing the correct method to save a TensorFlow model in SavedModel format

You have a trained TensorFlow Keras model and want to save it in the SavedModel format for later use in TensorFlow Serving. Which method should you use?

A tf.saved_model.save(model, 'path_to_dir')

B model.save('path_to_dir', save_format='tf')

C model.save_weights('path_to_dir')

D tf.keras.models.save_model(model, 'path_to_dir', save_format='h5')

💡 Hint

Think about the recommended way to save a full Keras model including architecture and weights in SavedModel format.

Using model.save('path', save_format='tf') saves the entire model in SavedModel format, including architecture, weights, and optimizer state. The tf.saved_model.save function is lower level and requires a concrete function. Saving weights only does not save the architecture. Saving with save_format='h5' saves in HDF5 format, not SavedModel.

❓ Hyperparameter

advanced

Effect of including signatures when saving a SavedModel

When saving a TensorFlow model using tf.saved_model.save, what is the effect of specifying the signatures argument?

A It automatically tunes hyperparameters during saving for better performance.

B It compresses the SavedModel files to reduce disk space usage.

C It defines the input and output signatures for serving, enabling TensorFlow Serving to know how to call the model.

D It encrypts the SavedModel to protect the model weights.

💡 Hint

Think about why TensorFlow Serving needs information about inputs and outputs.

The signatures argument specifies the input and output formats (signatures) of the model functions. This helps TensorFlow Serving or other tools know how to call the model correctly.

🔧 Debug

expert

Diagnosing a loading error with a SavedModel

You saved a TensorFlow model using model.save('my_model'). Later, when you try to load it with tf.keras.models.load_model('my_model'), you get the error:

ValueError: No model found in config.

What is the most likely cause of this error?

A The model contains custom layers or objects but they were not provided during loading.

B The model was saved in HDF5 format but loaded as SavedModel format.

C The TensorFlow version used to load is older than the version used to save.

D The SavedModel directory is missing the 'variables' folder.

💡 Hint

Consider what happens if the model has custom components that TensorFlow does not know about by default.

If the model has custom layers or functions, you must provide them via the custom_objects argument when loading. Otherwise, TensorFlow cannot reconstruct the model architecture and raises this error.