None Types and Operations in feed_dict

[dmclark17]

Hello,

I am attempting to use the InteractionNetwork with no global attributes. As mentioned in #22, I have set global_model_fn=None for the GraphIndependent block for encoding and decoding. My input placeholder graph tuple looks like this

GraphsTuple(
nodes=<tf.Tensor 'input_graph_placeholder/nodes:0' shape=(?, 6) dtype=float64>, 
edges=<tf.Tensor 'input_graph_placeholder/edges:0' shape=(?, 2) dtype=float64>,
receivers=<tf.Tensor 'input_graph_placeholder/receivers:0' shape=(?,) dtype=int32>, 
senders=<tf.Tensor 'input_graph_placeholder/senders:0' shape=(?,) dtype=int32>, 
globals=None, 
n_node=<tf.Tensor 'input_graph_placeholder/n_node:0' shape=(?,) dtype=int32>, 
n_edge=<tf.Tensor 'input_graph_placeholder/n_edge:0' shape=(?,) dtype=int32>
)

As expected globals = None, I then use the make_all_runnable_in_session as defined in the demos. This yields

GraphsTuple(
nodes=<tf.Tensor 'input_graph_placeholder/nodes:0' shape=(?, 6) dtype=float64>,
edges=<tf.Tensor 'input_graph_placeholder/edges:0' shape=(?, 2) dtype=float64>, 
receivers=<tf.Tensor 'input_graph_placeholder/receivers:0' shape=(?,) dtype=int32>,
senders=<tf.Tensor 'input_graph_placeholder/senders:0' shape=(?,) dtype=int32>, 
globals=<tf.Operation 'make_graph_runnable_in_session/NoOp' type=NoOp>,
n_node=<tf.Tensor 'input_graph_placeholder/n_node:0' shape=(?,) dtype=int32>, 
n_edge=<tf.Tensor 'input_graph_placeholder/n_edge:0' shape=(?,) dtype=int32>
)

As expected, this converts the None into a Tensorflow no_op. But when I try to run the session, I run into this error

TypeError                                 Traceback (most recent call last)
<ipython-input-6-ca8c0124fd42> in <module>()
     21       "loss": loss_op_tr,
     22       "outputs": output_ops_tr
---> 23   }, feed_dict=feed_dict)
     24 
     25 

/home/david/.local/lib/python2.7/site-packages/tensorflow/python/client/session.pyc in run(self, fetches, feed_dict, options, run_metadata)
    927     try:
    928       result = self._run(None, fetches, feed_dict, options_ptr,
--> 929                          run_metadata_ptr)
    930       if run_metadata:
    931         proto_data = tf_session.TF_GetBuffer(run_metadata_ptr)

/home/david/.local/lib/python2.7/site-packages/tensorflow/python/client/session.pyc in _run(self, handle, fetches, feed_dict, options, run_metadata)
   1093           except Exception as e:
   1094             raise TypeError(
-> 1095                 'Cannot interpret feed_dict key as Tensor: ' + e.args[0])
   1096 
   1097           if isinstance(subfeed_val, ops.Tensor):

TypeError: Cannot interpret feed_dict key as Tensor: Can not convert a Operation into a Tensor.

The TensorFlow documentation seems to indicate Operations cannot be keys of feed_dict. Is this how None types are suppose to be handled?

I am running Tensorflow version 1.12.0 with python 2.7 on Ubuntu 18.04

Thank you!

[vbapst]

Hi David,

make_all_runnable_in_session is meant to be used for the output of the graph, to convert GraphsTuple containing possibly None fields into runnable tensors.
In your case, you should do
feed_dict = utils_tf.get_feed_dict(placeholders, input_graph)
where placeholders is the GraphsTuple you correctly constructed in your first snippet.

[dmclark17]

Great! Thank you!

I am now running:

• make_all_runnable_in_session on output_ops_tr and output_ops_tr
• utils_tf.get_feed_dict(placeholders, input_graph) on my inputs and targets

And it is working for None fields.