Introduction

The TensorFlow error ValueError: No variables to save from means the saver or checkpoint logic cannot find any trainable or tracked variables at the moment you try to save. This usually happens because the model has not been built yet, variables were created outside the tracked object graph, or the code is mixing old TensorFlow 1 saving patterns with newer TensorFlow 2 APIs. The fix is to make sure actual variables exist and are attached to the object you are saving.

Common Cause in TensorFlow 2: Model Not Built Yet

In TensorFlow 2 and Keras, variables are often created lazily on the first call.

1import tensorflow as tf
2
3model = tf.keras.Sequential([
4    tf.keras.layers.Dense(16, activation="relu"),
5    tf.keras.layers.Dense(1),
6])
7
8print(model.variables)  # often empty before build or first call

If you try to save immediately, there may be nothing to save.

Build the model first:

model.build((None, 8))
print(len(model.variables))
model.save_weights("weights.ckpt")

You can also create variables by calling the model once with sample input.

Saving After a Forward Pass

Another valid pattern is to force variable creation by running data through the model.

1import tensorflow as tf
2
3model = tf.keras.Sequential([
4    tf.keras.layers.Dense(16, activation="relu"),
5    tf.keras.layers.Dense(1),
6])
7
8dummy = tf.random.uniform((4, 8))
9_ = model(dummy)
10
11print(len(model.variables))
12model.save_weights("weights.ckpt")

This is often the simplest fix when a lazily built model triggers the error.

TensorFlow 1 Saver Requires Graph Variables

In TensorFlow 1 style code, tf.train.Saver() expects variables in the graph collection. If you create no variables, or you are in the wrong graph context, the saver finds nothing.

1import tensorflow.compat.v1 as tf
2tf.disable_eager_execution()
3
4w = tf.Variable(3.0, name="weight")
5saver = tf.train.Saver()
6
7with tf.Session() as sess:
8    sess.run(tf.global_variables_initializer())
9    saver.save(sess, "model.ckpt")

If w were missing, Saver would fail because there would be no variables in the graph to checkpoint.

Variables Must Be Tracked by the Saved Object

With tf.train.Checkpoint, TensorFlow only saves tracked objects and variables.

Correct:

1import tensorflow as tf
2
3class MyModule(tf.Module):
4    def __init__(self):
5        super().__init__()
6        self.w = tf.Variable(2.0)
7
8module = MyModule()
9ckpt = tf.train.Checkpoint(module=module)
10ckpt.save("ckpts/ckpt")

If you create variables in local scope and never attach them to the module or model, checkpointing may not find them.

Mixed API Usage Causes Confusion

One common source of this error is mixing TensorFlow 1 graph-style save logic with TensorFlow 2 eager execution assumptions. If you are using modern Keras models, prefer:

• 'model.save()'
• 'model.save_weights()'
• 'tf.train.Checkpoint'

Do not reach for tf.train.Saver() unless you are explicitly working in TensorFlow 1 compatibility mode.

Inspect What TensorFlow Thinks Exists

Before saving, print the variables you expect TensorFlow to track.

for v in model.variables:
    print(v.name, v.shape)

For checkpoint objects:

print(model.trainable_variables)

If these are empty, the save failure is expected and the issue is earlier in model construction.

Practical Debugging Flow

When this error appears:

1. Check whether the model has been built.
2. Print tracked variables.
3. Confirm you are using one TensorFlow API style consistently.
4. Verify variables are attached to the saved model or module.

This usually finds the problem faster than experimenting with different save calls.

Common Pitfalls

• Saving a Keras model before it has created any variables.
• Creating variables in local scope and not attaching them to a tracked object.
• Mixing tf.train.Saver() with TensorFlow 2 eager-style code.
• Assuming model definition automatically creates variables without build or first call.
• Debugging the saver instead of inspecting whether any variables exist first.

Summary

• The error means TensorFlow cannot find tracked variables at save time.
• In TensorFlow 2, build the model or run one forward pass before saving.
• In TensorFlow 1, Saver needs graph variables to exist in the active graph.
• Prefer modern save APIs for modern Keras and TensorFlow code.
• Always inspect tracked variables before blaming the checkpoint mechanism.