Introduction

The error AttributeError: module 'tensorflow' has no attribute 'GraphDef' occurs when running TensorFlow 1.x code on TensorFlow 2.x. In TF 1.x, tf.GraphDef was a commonly used class for representing computational graphs as protocol buffer objects. In TF 2.x, this and many other attributes were reorganized under tf.compat.v1 or moved to different modules. This error is one of the most frequent migration issues when upgrading from TF 1.x to TF 2.x.

The Error

1import tensorflow as tf
2
3# TF 1.x code that fails in TF 2.x
4graph_def = tf.GraphDef()
5# AttributeError: module 'tensorflow' has no attribute 'GraphDef'

Fix 1: Use tf.compat.v1.GraphDef

The quickest fix — access the class through the compatibility module:

1import tensorflow as tf
2
3# TF 2.x compatible
4graph_def = tf.compat.v1.GraphDef()
5
6# Load a frozen graph
7with open('frozen_model.pb', 'rb') as f:
8    graph_def.ParseFromString(f.read())

Fix 2: Use tf.Graph().as_graph_def()

In TF 2.x, GraphDef is available as a return type from graph operations:

1import tensorflow as tf
2
3# Get GraphDef from an existing graph
4@tf.function
5def my_func(x):
6    return x * 2 + 1
7
8# Get the concrete function's graph
9concrete = my_func.get_concrete_function(tf.TensorSpec([None], tf.float32))
10graph_def = concrete.graph.as_graph_def()
11
12print(type(graph_def))  # <class 'tensorflow.core.framework.graph_pb2.GraphDef'>
13print(len(graph_def.node))  # Number of operations in the graph

Fix 3: Import from the Proto Module Directly

1from tensorflow.core.framework import graph_pb2
2
3graph_def = graph_pb2.GraphDef()
4
5# Parse a saved graph
6with open('model.pb', 'rb') as f:
7    graph_def.ParseFromString(f.read())
8
9# Inspect nodes
10for node in graph_def.node:
11    print(f"{node.name}: {node.op}")

Common TF 1.x → 2.x Attribute Migrations

Loading Frozen Graphs in TF 2.x

The common use case for GraphDef is loading frozen graphs (.pb files) from TF 1.x:

1import tensorflow as tf
2
3def load_frozen_graph(pb_file):
4    """Load a TF 1.x frozen graph in TF 2.x."""
5    with tf.io.gfile.GFile(pb_file, 'rb') as f:
6        graph_def = tf.compat.v1.GraphDef()
7        graph_def.ParseFromString(f.read())
8
9    # Wrap in a tf.function for TF 2.x usage
10    def model_fn(input_tensor):
11        output = tf.graph_util.import_graph_def(
12            graph_def,
13            input_map={'input:0': input_tensor},
14            return_elements=['output:0']
15        )
16        return output[0]
17
18    return model_fn
19
20# Usage
21model = load_frozen_graph('frozen_model.pb')
22result = model(tf.constant([[1.0, 2.0, 3.0]]))

Convert Frozen Graph to SavedModel

For a more permanent fix, convert the frozen graph to TF 2.x SavedModel format:

1import tensorflow as tf
2
3# Load the frozen graph
4graph_def = tf.compat.v1.GraphDef()
5with open('frozen_model.pb', 'rb') as f:
6    graph_def.ParseFromString(f.read())
7
8# Create a function from the graph
9wrapped = tf.compat.v1.wrap_function(
10    lambda: tf.compat.v1.import_graph_def(graph_def, name=''),
11    signature=[]
12)
13
14# Save as TF 2.x SavedModel
15tf.saved_model.save(wrapped, 'saved_model/')
16
17# Now load with standard TF 2.x API
18model = tf.saved_model.load('saved_model/')

Using tf.compat.v1 Broadly

For large codebases with many TF 1.x patterns, you can disable TF 2.x behavior entirely:

1import tensorflow.compat.v1 as tf
2tf.disable_v2_behavior()
3
4# Now TF 1.x code works as-is
5graph_def = tf.GraphDef()
6sess = tf.Session()
7# etc.

This is a temporary solution for migration. The long-term fix is to rewrite code using TF 2.x APIs.

Automatic Migration Script

TensorFlow provides a script to automatically update TF 1.x code:

1# Install the upgrade tool
2pip install tensorflow
3
4# Upgrade a single file
5tf_upgrade_v2 --infile old_code.py --outfile new_code.py
6
7# Upgrade an entire directory
8tf_upgrade_v2 --intree project/ --outtree project_v2/ --reportfile report.txt

The script converts tf.GraphDef → tf.compat.v1.GraphDef, tf.Session → tf.compat.v1.Session, and hundreds of other API changes. Review the generated report.txt for changes that need manual attention.

Common Pitfalls

• Using tf.compat.v1 as a permanent solution: tf.compat.v1 is a migration bridge, not a long-term API. TensorFlow may eventually deprecate it. Plan to migrate to native TF 2.x APIs (eager execution, tf.function, Keras).
• Mixed TF 1.x and 2.x code: Calling tf.compat.v1.disable_v2_behavior() affects the entire process. You cannot use TF 2.x features (like eager execution) after disabling v2 behavior. Choose one approach per script.
• Frozen graph input/output names: When loading frozen graphs, you must know the exact input and output tensor names (e.g., 'input:0', 'output:0'). Use graph_def.node to list all operations if you do not know the names.
• tf_upgrade_v2 is not perfect: The automatic migration script handles API renames but does not restructure code for eager execution or replace Session.run() patterns. Manual refactoring is still needed for idiomatic TF 2.x code.
• Version pinning: If migrating is not feasible, pin TensorFlow 1.x in your requirements: tensorflow==1.15.5. But note that TF 1.x no longer receives security updates.

Summary

• Replace tf.GraphDef with tf.compat.v1.GraphDef() for the quickest fix
• Import from tensorflow.core.framework.graph_pb2 for direct proto access
• Use tf_upgrade_v2 to automatically migrate TF 1.x code to TF 2.x compatible syntax
• Convert frozen graphs (.pb) to SavedModel format for native TF 2.x compatibility
• Plan to migrate from tf.compat.v1 to native TF 2.x APIs for long-term support