Introduction

In TensorFlow, a checkpoint and a GraphDef .pb file solve different problems even though both are associated with saved models. A checkpoint is primarily about restoring variables so you can resume or reuse model state, while a GraphDef .pb file is a serialized computation graph, often used for inference-oriented export in older TensorFlow workflows.

What a Checkpoint Contains

A TensorFlow checkpoint stores variable values and the information needed to restore those variables into a compatible object graph. In practical terms, checkpoints are for state restoration.

Typical uses include:

• continuing interrupted training
• loading pre-trained weights into the same model architecture
• restoring optimizer state when the checkpoint tracks it

A minimal TensorFlow 2 example looks like this.

1import tensorflow as tf
2
3weight = tf.Variable(3.0)
4ckpt = tf.train.Checkpoint(weight=weight)
5path = ckpt.save("/tmp/demo_ckpt")
6
7weight.assign(0.0)
8ckpt.restore(path).assert_consumed()
9print(weight.numpy())

After restoration, weight returns to the saved value.

The important point is that a checkpoint does not stand alone as a full deployable inference artifact. It assumes you know what variables should exist and where they belong.

What a GraphDef .pb File Contains

A GraphDef .pb file is a serialized protobuf representation of a TensorFlow graph. In older TensorFlow 1.x workflows, it was common to freeze a graph so that variable values became constants embedded into the graph, producing a single inference-oriented .pb file.

A small TensorFlow 2 example can still produce a frozen GraphDef.

1import tensorflow as tf
2from tensorflow.python.framework.convert_to_constants import convert_variables_to_constants_v2
3
4class AddOne(tf.Module):
5    @tf.function(input_signature=[tf.TensorSpec(shape=None, dtype=tf.float32)])
6    def __call__(self, x):
7        return x + 1.0
8
9module = AddOne()
10concrete = module.__call__.get_concrete_function()
11frozen = convert_variables_to_constants_v2(concrete)
12
13tf.io.write_graph(frozen.graph, "/tmp", "add_one.pb", as_text=False)

That .pb file describes computation. In a frozen graph scenario, it may also contain the constant values needed for inference.

The Practical Difference

The simplest mental model is:

• checkpoint equals restorable state
• GraphDef .pb equals serialized graph structure

That difference drives how each format is used.

If you want to fine-tune a model, continue training, or swap weights into the same architecture, use a checkpoint.

If you want an older-style inference graph that can be loaded and executed as a frozen computation, a .pb graph is the relevant artifact.

Why Confusion Happens

Many pre-trained TensorFlow models used to ship as a mix of graph files, checkpoint files, label maps, and config files. People would see both a checkpoint and a .pb file in the same project and assume they were interchangeable.

They are not.

A checkpoint generally needs compatible code or an equivalent object structure to restore into. A frozen .pb graph is closer to a portable inference representation, but it is less convenient for further training because the variables may no longer exist as variables.

Legacy TensorFlow vs Current Practice

The checkpoint-versus-GraphDef distinction comes mostly from TensorFlow 1.x mental models. In current TensorFlow workflows, you will more often see:

• checkpoints for training state
• 'SavedModel for export and serving'
• '.keras for Keras-native saving workflows'

That means raw .pb questions are often legacy deployment questions rather than the default modern recommendation.

Still, understanding the distinction matters when you inherit older pre-trained model packages or need to convert legacy assets.

Common Pitfalls

The most common mistake is trying to resume training from a .pb graph as if it were a normal checkpoint.

Another mistake is assuming a checkpoint is self-sufficient for inference deployment. It is not, because the runtime still needs the compatible model structure.

A third issue is mixing TensorFlow 1.x terminology with TensorFlow 2 saving APIs without noticing that the preferred export formats changed.

Finally, do not assume that a file extension tells the whole story. The actual contents and the surrounding loading code matter.

Summary

• A checkpoint stores variable state for restoration and continued model use.
• A GraphDef .pb file stores a serialized computation graph, often for legacy inference export.
• Checkpoints are the right tool for resuming training or loading weights into a compatible model.
• Frozen .pb graphs are more aligned with inference than with ongoing training.
• In modern TensorFlow, SavedModel and .keras are often more relevant than raw GraphDef files.
• When working with pre-trained assets, identify whether you need state restoration or graph export before choosing the format.