Implementation:Tensorflow Serving SavedModelBuilder Export

Knowledge Sources	TensorFlow Serving SavedModelBuilder API
Domains	Model_Serialization, Deployment
Last Updated	2026-02-13 17:00 GMT

Overview

Wrapper for TensorFlow's SavedModelBuilder class to serialize trained models with signatures to the SavedModel directory format.

Description

SavedModelBuilder is TensorFlow's builder pattern API for constructing SavedModel artifacts. In TensorFlow Serving's MNIST example, it:

Creates the export directory at <export_dir>/<model_version>/
Adds the trained session's MetaGraph with serving tags and signature definitions
Writes the serialized protobuf and checkpoint files to disk

The builder validates that the export path does not already exist (to prevent accidental overwrites) and supports multiple MetaGraphs with different tag sets in a single SavedModel.

Usage

Use this after defining SignatureDefs and completing model training. The builder is the bridge between a live TensorFlow session and the on-disk format that TensorFlow Serving can load.

Code Reference

Source Location

Repository: tensorflow/serving
File: tensorflow_serving/example/mnist_saved_model.py
Lines: L99-153

Signature

# Constructor
tf.compat.v1.saved_model.builder.SavedModelBuilder(
    export_dir: str  # Path where SavedModel will be written
)

# Add MetaGraph with variables
builder.add_meta_graph_and_variables(
    sess: tf.Session,                    # Trained session with variables
    tags: list[str],                     # Tags identifying this MetaGraph (e.g., ["serve"])
    signature_def_map: dict[str, SignatureDef],  # Named signatures
    main_op: tf.Operation = None,        # Op to run on load (e.g., table initializer)
    strip_default_attrs: bool = False    # Remove default-valued attributes for forward compat
)

# Serialize to disk
builder.save() -> None

Import

import tensorflow as tf
# Access via: tf.compat.v1.saved_model.builder.SavedModelBuilder

I/O Contract

Inputs

Name	Type	Required	Description
export_dir	str	Yes	Full path including version subdirectory
sess	tf.Session	Yes	Trained TF session with variable values
tags	list[str]	Yes	Tags for MetaGraph (typically ["serve"])
signature_def_map	dict[str, SignatureDef]	Yes	Named signatures for serving
main_op	tf.Operation	No	Op to run when model is loaded
strip_default_attrs	bool	No	Default False; strip default attributes

Outputs

Name	Type	Description
saved_model.pb	File	Serialized MetaGraphDef protobuf
variables/	Directory	Checkpoint files with trained weights

Usage Examples

Export MNIST Model

import os
import tensorflow as tf

export_path = os.path.join('/tmp/mnist_model', str(1))  # version 1
builder = tf.compat.v1.saved_model.builder.SavedModelBuilder(export_path)

# Add MetaGraph with signatures
builder.add_meta_graph_and_variables(
    sess,
    [tf.compat.v1.saved_model.tag_constants.SERVING],
    signature_def_map={
        'predict_images': prediction_signature,
        tf.compat.v1.saved_model.signature_constants.DEFAULT_SERVING_SIGNATURE_DEF_KEY:
            classification_signature,
    },
    main_op=tf.compat.v1.tables_initializer(),
    strip_default_attrs=True
)

builder.save()
print(f'Model exported to {export_path}')

Command Line Usage

# Export model version 1
python tensorflow_serving/example/mnist_saved_model.py /tmp/mnist_model

# Export model version 2 (for version management)
python tensorflow_serving/example/mnist_saved_model.py \
    --model_version=2 \
    --training_iteration=2000 \
    /tmp/mnist_model

Related Pages

Page Connections

Double-click a node to navigate. Hold to expand connections.

Principle

Implementation

Heuristic

Environment