How do I get TensorFlow's 'import_graph_def' to return Tensors

Question

If I attempt to import a saved TensorFlow graph definition with

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

with gfile.FastGFile(FLAGS.model_save_dir.format(log_id) + '/graph.pb', 'rb') as f:
    graph_def = tf.GraphDef()
    graph_def.ParseFromString(f.read())
x, y, y_ = tf.import_graph_def(graph_def, 
                               return_elements=['data/inputs',
                                                'output/network_activation',
                                                'data/correct_outputs'],
                               name='')

the returned values are not Tensors as expected, but something else: instead, for example, of getting x as

Tensor("data/inputs:0", shape=(?, 784), dtype=float32)

I get

name: "data/inputs_1"
op: "Placeholder"
attr {
  key: "dtype"
  value {
    type: DT_FLOAT
  }
}
attr {
  key: "shape"
  value {
    shape {
    }
  }
}

That is, instead of getting the expected tensor x I get, x.op. This confuses me because the documentation seems to say I should get a Tensor (though there are a bunch of ors there that make it hard to understand).

How do I get tf.import_graph_def to return specific Tensors that I can then use (e.g. in feeding the loaded model, or running analyses)?

Answer 1

The names 'data/inputs', 'output/network_activation' and 'data/correct_outputs' are actually operation names. To get tf.import_graph_def() to return tf.Tensor objects, you should append the an output index to the operation name, which is typically ':0' for single-output operations:

x, y, y_ = tf.import_graph_def(graph_def, 
                               return_elements=['data/inputs:0',
                                                'output/network_activation:0',
                                                'data/correct_outputs:0'],
                               name='')