TF save/restore graph fails at tf.GraphDef.ParseFromString()

Question

Based on this converting-trained-tensorflow-model-to-protobuf I am trying to save/restore TF graph without success.

Here is saver:

with tf.Graph().as_default():
    variable_node = tf.Variable(1.0, name="variable_node")
    output_node = tf.mul(variable_node, 2.0, name="output_node")
    sess = tf.Session()
    init = tf.initialize_all_variables()
    sess.run(init)
    output = sess.run(output_node)
    tf.train.write_graph(sess.graph.as_graph_def(), summ_dir, 'model_00_g.pbtxt', as_text=True)
    #self.assertNear(2.0, output, 0.00001)
    saver = tf.train.Saver()
    saver.save(sess, saver_path)

which produces model_00_g.pbtxt with text graph description. Pretty much copy paste from freeze_graph_test.py.

Here is reader:

with tf.Session() as sess:

    with tf.Graph().as_default():
        graph_def = tf.GraphDef()
        graph_path = '/mnt/code/test_00/log/2016-02-11.22-37-46/model_00_g.pbtxt'
        with open(graph_path, "rb") as f:
            proto_b = f.read()
            #print proto_b   # -> I can see it
            graph_def.ParseFromString(proto_b) # no luck..
            _ = tf.import_graph_def(graph_def, name="")

    print sess.graph_def

which fails at graph_def.ParseFromString() with DecodeError: Tag had invalid wire type.

I am on docker container b.gcr.io/tensorflow/tensorflow:latest-devel in case it makes any difference.

score 19 · Accepted Answer · answered Feb 11 '16 at 23:31

19

The GraphDef.ParseFromString() method (and, in general, the ParseFromString() method on any Python protobuf wrapper) expects a string in the binary protocol buffer format. If you pass as_text=False to tf.train.write_graph(), then the file will be in the appropriate format.

Otherwise you can do the following to read the text-based format:

from google.protobuf import text_format
# ...
graph_def = tf.GraphDef()
text_format.Merge(proto_b, graph_def)

answered Feb 11 '16 at 23:31

mrry

125,488
26
399
400

thanks - these both way solve immediate problem. However main task is not here for me - i want to store graph+vars and load then for eval in other place. This [`tensorflow/tensorflow/python/tools/freeze_graph.py`](https://github.com/tensorflow/tensorflow/tree/00440e99ffb1ed1cfe4b4ea650e0c560838a6edc/tensorflow/python/tools) looks very much as what i need, but it is not in my docker image yet and usage isn't very clear wrt half of parameters - I think i will wait 0.7 for this and switch to other tasks for now. – rgr Feb 12 '16 at 20:26
getting the same kind of problem. Not able to solve, can you look into this https://stackoverflow.com/questions/71439124/google-protobuf-message-decodeerror-error-parsing-message-with-type-tensorflow – subbu Mar 14 '22 at 12:19

score 1 · Answer 2 · answered Feb 11 '16 at 23:28

1

ParseFromString needs binary serialized protocol buffer, for human-readable representation you need to use text_format.Merge as used here

answered Feb 11 '16 at 23:28

Yaroslav Bulatov

57,332
22
139
197

score 0 · Answer 3 · answered May 12 '17 at 16:44

I tried to load model via the java API which accepts only binary. But in the python where we uses contrib.Estimator produces textual model file format. I found a model file converter online, seems it's working fine. This might solve the original issue too (use the binary model loader) if you have an existing text format model file.

TF save/restore graph fails at tf.GraphDef.ParseFromString()

3 Answers3

Linked