TensorFlow SavedModel: Inspecting SavedModel Contents

TensorFlow’s SavedModel format is an essential factor in ensuring that your trained machine learning models can be reused, evaluated, and served without complexity. In this article, we will delve into TensorFlow's SavedModel, understand how to inspect its contents, and walk through practical code examples to achieve these goals.

Understanding TensorFlow SavedModel

The SavedModel format is TensorFlow's standard format for serializing models. SavedModels are versatile and can be deployed across different environments or serving platforms. Each SavedModel directory will contain all the necessary files to unpack its functionality for inference, evaluation, or further training.

SavedModel Directory Structure

A typical SavedModel directory contains a few critical components:

• saved_model.pb: Protocol buffer file containing a serialized description of the model, including its architecture and assets.
• Variables: A folder that holds variable data files and their checkpoint information.
• Assets: A folder to host any external files required by TensorFlow operations that the model references.

Inspecting a SavedModel

When you want to inspect the contents of a SavedModel, TensorFlow offers a saved_model_cli utility tool that’s a command line interface for SavedModel inspection without running a Python script.

saved_model_cli show --dir /path/to/saved_model --all

This command reveals detailed information about the MetaGraphDefs and SignatureDefs included within your SavedModel. Example output will look something like this:

MetaGraphDef with tag-set 'serve' contains the following SignatureDefs:

Inspecting SavedModel using Python

Beyond the CLI tool, inspecting a SavedModel with Python can provide interactive insights. You can use Python to inspect and manipulate SavedModel contents programmatically:

import tensorflow as tf

saved_model_path = '/path/to/saved_model'
saved_model = tf.saved_model.load(saved_model_path)

# Print available signatures
print(list(saved_model.signatures.keys()))

This basic code will load the SavedModel and print out the list of defined signatures, usually showcasing inputs required and outputs provided by the model—all of which can help determine if the model meets your feature requirements.

Loading and Using a SavedModel

After determining that the model meets your needs, loading and making predictions with the model is straightforward using TensorFlow’s saved_model.load() function. Let's look at a full example:

import tensorflow as tf

def perform_inference(saved_model_path, input_data):
    # Load the saved model
    model = tf.saved_model.load(saved_model_path)
    infer = model.signatures['serving_default']

    # Run inference
    result = infer(tf.constant(input_data))
    print(result)

# Example usage
sample_input = ... # Your input data for inference
demo_saved_model = '/path/to/saved_model'
perform_inference(demo_saved_model, sample_input)

This Python script loads a SavedModel and performs inference on given input data using the default signature.

Conclusion

Inspecting your SavedModel not only helps verify its integrity but also clarifies its capabilities in terms of supported inputs and outputs. Understanding the model's assets thanks to easy Python functions or even the command line tool can immensely streamline the development and deployment processes. With this approach, practitioners can seamlessly and effectively integrate machine learning models into larger applications.