Tensorflow freeze_graph script failing on model defined with Keras

Question

I'm attempting to export a model built and trained with Keras to a protobuffer that I can load in a C++ script (as in this example). I've generated a .pb file containing the model definition and a .ckpt file containing the checkpoint data. However, when I try to merge them into a single file using the freeze_graph script I get the error:

ValueError: Fetch argument 'save/restore_all' of 'save/restore_all' cannot be interpreted as a Tensor. ("The name 'save/restore_all' refers to an Operation not in the graph.")

I'm saving the model like this:

with tf.Session() as sess:
    model = nndetector.architecture.models.vgg19((3, 50, 50))
    model.load_weights('/srv/nn/weights/scratch-vgg19.h5')
    init_op = tf.initialize_all_variables()
    sess.run(init_op)
    graph_def = sess.graph.as_graph_def()
    tf.train.write_graph(graph_def=graph_def, logdir='.',   name='model.pb', as_text=False)
    saver = tf.train.Saver()
    saver.save(sess, 'model.ckpt')

nndetector.architecture.models.vgg19((3, 50, 50)) is simply a vgg19-like model defined in Keras.

I'm calling the freeze_graph script like this:

bazel-bin/tensorflow/python/tools/freeze_graph --input_graph=[path-to-model.pb] --input_checkpoint=[path-to-model.ckpt] --output_graph=[output-path] --output_node_names=sigmoid --input_binary=True

If I run the freeze_graph_test script everything works fine.

Does anyone know what I'm doing wrong?

Thanks.

Best regards

Philip

EDIT

I've tried printing tf.train.Saver().as_saver_def().restore_op_name which returns save/restore_all.

Additionally, I've tried a simple pure tensorflow example and still get the same error:

a = tf.Variable(tf.constant(1), name='a')
b = tf.Variable(tf.constant(2), name='b')
add = tf.add(a, b, 'sum')

with tf.Session() as sess:
sess.run(tf.initialize_all_variables())
tf.train.write_graph(graph_def=sess.graph.as_graph_def(), logdir='.',     name='simple_as_binary.pb', as_text=False)
tf.train.Saver().save(sess, 'simple.ckpt')

And I'm actually also unable to restore the graph in python. Using the following code throws ValueError: No variables to save if I execute it separately from saving the graph (that is, if I both save and restore the model in the same script, everything works fine).

with gfile.FastGFile('simple_as_binary.pb') as f:
    graph_def = tf.GraphDef()
    graph_def.ParseFromString(f.read())

with tf.Session() as sess:
    tf.import_graph_def(graph_def)
    saver = tf.train.Saver()
    saver.restore(sess, 'simple.ckpt')

I'm not sure if the two problems are related, or if I'm simply not restoring the model correctly in python.

Answer 1

The problem is the order of these two lines in your original program:

tf.train.write_graph(graph_def=sess.graph.as_graph_def(), logdir='.',     name='simple_as_binary.pb', as_text=False)
tf.train.Saver().save(sess, 'simple.ckpt')

Calling tf.train.Saver() adds a set of nodes to the graph, including one called "save/restore_all". However, this program calls it after writing out the graph, so the file you pass to freeze_graph.py doesn't contain those nodes, which are necessary for doing the rewriting.

Reversing the two lines should make the script work as intended:

tf.train.Saver().save(sess, 'simple.ckpt')
tf.train.write_graph(graph_def=sess.graph.as_graph_def(), logdir='.',     name='simple_as_binary.pb', as_text=False)