Is there a way to save/restore the graph(s) of a tf.function

Question

I use in my project large graphs that can take more than 5min of compilation time. I decorated the large intermediate parts using a tf.function(reduce_retracing=True). I also heavily use tf.experimental.ExtensionType instances.

My question is, would it be possible to cache the graphs compiled by these functions, and restore those graphs on a different execution? In the case where multiple concrete functions have been retraced, i would like to save them all.

I precise that i am not interested in model saving, since it saves weights not graphs, and needs recompiling anyways.

Moreover, i have tried the following code:

import tensorflow as tf
@tf.function(input_signature=[tf.TensorSpec(shape=None, dtype=tf.float32)])
def f(x):
    print('retrace')
    return x + 1.
f(tf.constant(0.)) # prints 'retrace'
exported = tf.train.Checkpoint(f=f)
tf.saved_model.save(exported, 'tf_cache/')
imported = tf.saved_model.load('tf_cache/') # prints 'retrace'
imported.f(tf.constant(0.))

The issue is that i get a retrace when loading the saved checkpoint. Moreover, when f uses tf.experimental.ExtensionType inputs/outputs, tf.save_model.save raises a ValueError.

Thanks in advance for any answer.

Mean Coder · Accepted Answer · 2025-08-02 10:39:01Z

0

Your code works fine with me TF2.12.0 however it does not print anything. Use tf.print instead of print if you want print statements in your tf.function.

answered Aug 2 at 10:39

Mean Coder

3041 silver badge12 bronze badges

Sign up to request clarification or add additional context in comments.

Collectives™ on Stack Overflow

Is there a way to save/restore the graph(s) of a tf.function

1 Answer 1

Comments

Your Answer

Hot Network Questions

Collectives™ on Stack Overflow

1 Answer 1

Comments

Your Answer

Sign up or log in

Post as a guest

Related