Script that reproduce issues with tf.train.write_graph() and graph_utils.convert_to_constant_graph()

demo.py

from __future__ import division
from __future__ import print_function
from __future__ import unicode_literals

import numpy as np
import os
import tensorflow as tf
from tensorflow.python.client import graph_util

DATA_PLACEHOLDER_NAME = 'data'
IS_TRAIN_PHASE_PLACEHOLDER_NAME = 'is_train_phase'
RESULT_NODE_NAME = 'result'
DATA_BATCH_SHAPE = (5, 5)


def train_and_save_graph(checkpoints_dir):
    with tf.Graph().as_default():
        data_placeholder = tf.placeholder(dtype=tf.float32, shape=DATA_BATCH_SHAPE, name=DATA_PLACEHOLDER_NAME)
        is_train_phase_placeholder = tf.placeholder(dtype=tf.bool, name=IS_TRAIN_PHASE_PLACEHOLDER_NAME)
        
        # The actual computations are meaningless, but the pattern are similar to 
        # calculation of batch mean/variance for batch normalization
        batch_mean, batch_var = tf.nn.moments(data_placeholder, [0], name='moments')
        ema = tf.train.ExponentialMovingAverage(decay=0.9)
        ema_apply_op = ema.apply([batch_mean, batch_var])
        ema_mean = ema.average(batch_mean)
        ema_var = ema.average(batch_var)

        def mean_var_with_update():
            with tf.control_dependencies([ema_apply_op]):
                return tf.identity(batch_mean), tf.identity(batch_var)

        mean, var = tf.cond(is_train_phase_placeholder, mean_var_with_update, lambda: (ema_mean, ema_var))

        normalized_data = (data_placeholder - mean) / var

        result_op = tf.reduce_sum(normalized_data ** 2, name=RESULT_NODE_NAME)

        saver = tf.train.Saver(tf.all_variables())

        init_op = tf.initialize_all_variables()

        with tf.Session() as session:

            session.run(init_op)

            current_step = None
            for i in range(100):
                data = np.random.randn(*DATA_BATCH_SHAPE) * np.random.uniform(0.1, 10)
                result = session.run([result_op], feed_dict={data_placeholder: data, is_train_phase_placeholder: True})
                print('Step {:d}. result: {}'.format(i, result))
                current_step = i

            path_to_checkpoint = os.path.join(checkpoints_dir, 'model.ckpt')
            try:
                os.mkdir(checkpoints_dir)
            except OSError:
                pass

            saver.save(session, path_to_checkpoint, global_step=current_step)


def convert_to_constant_graph(checkpoints_dir, constant_graph_dir, constant_graph_file):
    with tf.Graph().as_default():
        with tf.Session() as session:
            checkpoint = tf.train.get_checkpoint_state(checkpoints_dir)
            if checkpoint and checkpoint.model_checkpoint_path:
                saver = tf.train.import_meta_graph(checkpoint.model_checkpoint_path + '.meta')
            else:
                raise ValueError('No checkpoint file found')
            saver.restore(session, checkpoint.model_checkpoint_path)
            constant_graph = graph_util.convert_variables_to_constants(session, session.graph_def, [RESULT_NODE_NAME])
            tf.train.write_graph(constant_graph, constant_graph_dir, constant_graph_file, as_text=False)


def use_saved_graph(constant_graph_dir, constant_graph_file):
    with tf.Graph().as_default():
        graph_def = tf.GraphDef()
        path_to_graph_file = os.path.join(constant_graph_dir, constant_graph_file)

        with open(path_to_graph_file, 'rb') as graph_file:
            graph_def.ParseFromString(graph_file.read())
            tf.import_graph_def(graph_def)

        with tf.Session() as session:
            result_op = session.graph.get_tensor_by_name(RESULT_NODE_NAME + ':0')
            data_placeholder = session.graph.get_tensor_by_name(DATA_PLACEHOLDER_NAME + ':0')
            is_train_phase_placeholder = session.get_tensor_by_name(DATA_PLACEHOLDER_NAME + ':0')

            data = 4.5 * np.random.randn(*DATA_BATCH_SHAPE)
            result = session.run([result_op], feed_dict={data_placeholder: data, is_train_phase_placeholder: False})
            print('Result is ', result)


checkpoint_dir = 'checkpoints'
graph_dir = 'graphs'
graph_file_name = 'constant_graph.pb'


train_and_save_graph(checkpoint_dir)
convert_to_constant_graph(checkpoint_dir, graph_dir, graph_file_name)
use_saved_graph(graph_dir, graph_file_name)

trace

Traceback (most recent call last):
  File "scripts/demo.py", line 99, in <module>
    use_saved_graph(graph_dir, graph_file_name)
  File "scripts/demo.py", line 80, in use_saved_graph
    tf.import_graph_def(graph_def)
  File "/home/ingwar/virtualenvs/upwork_data_science/local/lib/python2.7/site-packages/tensorflow/python/framework/importer.py", line 320, in import_graph_def
    node, 'Input tensor %r %s' % (input_name, te)))
ValueError: graph_def is invalid at node u'ExponentialMovingAverage/AssignMovingAvg': Input tensor 'moments/moments_1/mean/ExponentialMovingAverage:0' Cannot convert a tensor of type float32 to an input of type float32_ref.

Hi, I encountered exactly the same error. Have you worked around this issue?

Did you ever find a solution to this issue? Also hitting it when using ExponentialMovingAverage.

Thanks!

Did you find a solution? I also faced the same issue 😢