Convert .pb file into frozen .pb file

[nwo...@ebay.com]

Hello,

I have been struggling with this conversion. I tried the following:

import tensorflow as tf
import sys
from tensorflow.python.platform import gfile

from tensorflow.core.protobuf import saved_model_pb2
from tensorflow.python.util import compat

with tf.Session() as sess:
    model_filename ='saved_model.pb' # binary .pb file
    with gfile.FastGFile(model_filename, 'rb') as f:

        data = compat.as_bytes(f.read()) # reads binary
        sm = saved_model_pb2.SavedModel() 
        print(sm)
        sm.ParseFromString(data) # parses through the file
        print(sm)
        if 1 != len(sm.meta_graphs):
            print('More than one graph found. Not sure which to write')
            sys.exit(1)

        g_in = tf.import_graph_def(sm.meta_graphs[0].graph_def)
        output_graph = "frozen_grapha.pb"

        # This not working fully  
        output_graph_def = tf.graph_util.convert_variables_to_constants(
            sess, # The session is used to retrieve the weights
            tf.get_default_graph().as_graph_def(), # The graph_def is used to retrieve the nodes 
            ['softmax']# The output node names are used to select the usefull nodes
        ) 

        # Finally we serialize and dump the output graph to the filesystem
        with tf.gfile.GFile(output_graph, "wb") as f:
            f.write(output_graph_def.SerializeToString())
        print("%d ops in the final graph." % len(output_graph_def.node))
        print(g_in)
LOGDIR='.'
train_writer = tf.summary.FileWriter(LOGDIR)
train_writer.add_graph(sess.graph)

Without the output graph definition, it creates a event.out file that is readable by tensorboard. 

Could someone tell me what is wrong?

Best,

Nyla 

[Andrew Selle]

you can use sess.graph_def instead of tf.get_default_graph().as_graph_def(). I don't see anything wrong. What exactly is going wrong?

[nwo...@ebay.com]

Hello!

The code was not running within Jupyter notebooks, but kinda works from my computer when I replaced ['softmax'] with output_nodes = [n.name for n in tf.get_default_graph().as_graph_def().node], but now I have another bug: tensorflow.python.framework.errors_impl.FailedPreconditionError:Attempting to use uninitialized value . I found someone that fixed this, because they were constructing the Saver before I made my session. Is that what is going on within the code? It does not appear evident to me.   

Could there be something wrong with the .pb file I have?

Thank you!

On Monday, August 13, 2018 at 10:48:42 AM UTC-7, nwo...@ebay.com wrote:

[Andrew Selle]

You are using only the meta graph out of the saved model. This means you aren't loading  the checkpoint.

Look at how tflite does it

https://github.com/tensorflow/tensorflow/blob/37bfe7a9f290c267ff7a804038fb6c8979975dee/tensorflow/contrib/lite/python/lite.py#L259

And if you have further problems create a github bug with complete information and post it here. This discussion should be moved there.

-A

[nwo...@ebay.com]

Hey! 

I posted as a bug here: https://github.com/tensorflow/tensorflow/issues/21585

I would really appreciate some help. I tried initializing the variables, but couldn't do it.

Thank you