Minimal code to load a trained TensorFlow model from a checkpoint and export it with SavedModelBuilder

export_tf_model.py

import os
import tensorflow as tf

trained_checkpoint_prefix = 'checkpoints/dev'
export_dir = os.path.join('models', '0') # IMPORTANT: each model folder must be named '0', '1', ... Otherwise it will fail!

loaded_graph = tf.Graph()
with tf.Session(graph=loaded_graph) as sess:
    # Restore from checkpoint
    loader = tf.train.import_meta_graph(trained_checkpoint_prefix + '.meta')
    loader.restore(sess, trained_checkpoint_prefix)
    
    # Export checkpoint to SavedModel
    builder = tf.saved_model.builder.SavedModelBuilder(export_dir)
    builder.add_meta_graph_and_variables(sess,
                                         [tf.saved_model.tag_constants.TRAINING, tf.saved_model.tag_constants.SERVING],
                                         strip_default_attrs=True)
    builder.save()

Thanks !
Using compat for TensorFlow 2.0:

import os
import tensorflow as tf

# trained_checkpoint_prefix = 'checkpoints/dev'
trained_checkpoint_prefix = 'model.ckpt'
export_dir = os.path.join('models', '0') # IMPORTANT: each model folder must be named '0', '1', ... Otherwise it will fail!

loaded_graph = tf.Graph()
with tf.compat.v1.Session(graph=loaded_graph) as sess:
    # Restore from checkpoint
    loader = tf.compat.v1.train.import_meta_graph(trained_checkpoint_prefix + '.meta')
    loader.restore(sess, trained_checkpoint_prefix)
    
    # Export checkpoint to SavedModel
    builder = tf.compat.v1.saved_model.builder.SavedModelBuilder(export_dir)
    builder.add_meta_graph_and_variables(sess, ["train", "serve"], strip_default_attrs=True)
    builder.save()

I am getting a ValueError: Atleast 2 variables have the same name

Do you have an idea on how to solve this ?

signature_def_map is lost in this way

signature_def_map is lost in this way

Yes, and this is needed for serving. How to keep this parameter?

Any resolution to this? @Litchilitchy

@Aroueterra You have to add the input and output by yourself.

According to Tensorflow, it is like you must have an entry point. Although it make sense that the model to save already include this, like if a Keras model we could just use model.inputs, model.outputs.

Is it possible to directly get it from the code from .ckpt and .meta?