Request for Multi Graph composition

Peng Yu

unread,

Jul 5, 2018, 9:54:24 AM7/5/18

to dis...@tensorflow.org

Hi,

Let's say we have two graph/model being defined beforehand, and we want to connect the two graph into one. the output of graph_a redirect into graph_b.

But since graph_a and graph_b was being defined as fully functionally graph. Which with placeholder there is no way to connect them, except for eval one after the other in memory, which would lose the gradient propagation benefit

Is there any thoughts on that?

Martin Wicke

unread,

Jul 5, 2018, 10:50:14 AM7/5/18

to Peng Yu, Discuss

If you load both into the same Graph object, you can pass an output Tensor in a feed_dict. That will implicitly and temporarily connect the two graphs.

--
You received this message because you are subscribed to the Google Groups "Discuss" group.
To unsubscribe from this group and stop receiving emails from it, send an email to discuss+u...@tensorflow.org.
To post to this group, send email to dis...@tensorflow.org.
To view this discussion on the web visit https://groups.google.com/a/tensorflow.org/d/msgid/discuss/CAOKWFaJPNXuF0a7WYk2Y0QV9T%3DCfEKAGKjs7UJbQv3qQwviCBA%40mail.gmail.com.

Peng Yu

unread,

Jul 5, 2018, 4:13:29 PM7/5/18

to Martin Wicke, Discuss

it doesn't work as if i pass the tensor as value.

```

In [10]: import tensorflow as tf

In [11]: with tf.Session() as sess:

...: with tf.Graph().as_default() as g:

...: x = tf.placeholder(tf.int32, [None])

...: a = x + 1

...: y = tf.placeholder(tf.int32, [None])

...: b = y + 1

...: print(b.eval({y:a, x: [1]}))

...:

---------------------------------------------------------------------------

ValueError Traceback (most recent call last)

<ipython-input-11-551942816e5d> in <module>()

5 y = tf.placeholder(tf.int32, [None])

6 b = y + 1

----> 7 print(b.eval({y:a, x: [1]}))

8

/Users/pengyu/.pyenv/versions/2.7.13/lib/python2.7/site-packages/tensorflow/python/framework/ops.pyc in eval(self, feed_dict, session)

708

709 """

--> 710 return _eval_using_default_session(self, feed_dict, self.graph, session)

711

712

/Users/pengyu/.pyenv/versions/2.7.13/lib/python2.7/site-packages/tensorflow/python/framework/ops.pyc in _eval_using_default_session(tensors, feed_dict, graph, session)

5169 "`eval(session=sess)`")

5170 if session.graph is not graph:

-> 5171 raise ValueError("Cannot use the default session to evaluate tensor: "

5172 "the tensor's graph is different from the session's "

5173 "graph. Pass an explicit session to "

ValueError: Cannot use the default session to evaluate tensor: the tensor's graph is different from the session's graph. Pass an explicit session to `eval(session=sess)`.

```

On Thu, Jul 5, 2018 at 10:49 AM, Martin Wicke <wi...@google.com> wrote:

If you load both into the same Graph object, you can pass an output Tensor in a feed_dict. That will implicitly and temporarily connect the two graphs.

On Thu, Jul 5, 2018, 06:54 'Peng Yu' via Discuss <dis...@tensorflow.org> wrote:

Hi,

Let's say we have two graph/model being defined beforehand, and we want to connect the two graph into one. the output of graph_a redirect into graph_b.

But since graph_a and graph_b was being defined as fully functionally graph. Which with placeholder there is no way to connect them, except for eval one after the other in memory, which would lose the gradient propagation benefit

Is there any thoughts on that?

--
You received this message because you are subscribed to the Google Groups "Discuss" group.

To unsubscribe from this group and stop receiving emails from it, send an email to discuss+unsubscribe@tensorflow.org.

Peng Yu

unread,

Jul 5, 2018, 4:15:41 PM7/5/18

to Martin Wicke, Discuss

```

In [13]: with tf.Session() as sess:

...: with tf.Graph().as_default() as g:

...: x = tf.placeholder(tf.int32, [None])

...: a = x + 1

...: y = tf.placeholder(tf.int32, [None])

...: b = y + 1

...: print(sess.run(b, feed_dict={y:a, x:[1]}))

...:

---------------------------------------------------------------------------

RuntimeError Traceback (most recent call last)

<ipython-input-13-b28a8dc00625> in <module>()

5 y = tf.placeholder(tf.int32, [None])

6 b = y + 1

----> 7 print(sess.run(b, feed_dict={y:a, x:[1]}))

8

9

/Users/pengyu/.pyenv/versions/2.7.13/lib/python2.7/site-packages/tensorflow/python/client/session.pyc in run(self, fetches, feed_dict, options, run_metadata)

898 try:

899 result = self._run(None, fetches, feed_dict, options_ptr,

--> 900 run_metadata_ptr)

901 if run_metadata:

902 proto_data = tf_session.TF_GetBuffer(run_metadata_ptr)

/Users/pengyu/.pyenv/versions/2.7.13/lib/python2.7/site-packages/tensorflow/python/client/session.pyc in _run(self, handle, fetches, feed_dict, options, run_metadata)

1058 raise RuntimeError('Attempted to use a closed Session.')

1059 if self.graph.version == 0:

-> 1060 raise RuntimeError('The Session graph is empty. Add operations to the '

1061 'graph before calling run().')

1062

RuntimeError: The Session graph is empty. Add operations to the graph before calling run().

```

Martin Wicke

unread,

Jul 5, 2018, 4:54:51 PM7/5/18

to Peng Yu, Discuss

Don't to .eval, use session.run.

To unsubscribe from this group and stop receiving emails from it, send an email to discuss+u...@tensorflow.org.

Peng Yu

unread,

Jul 6, 2018, 10:01:36 AM7/6/18

to Martin Wicke, Discuss

yeah, i did that, check the latest reply ^ still it would complain

RuntimeError: The Session graph is empty. Add operations to the graph before calling run().

To unsubscribe from this group and stop receiving emails from it, send an email to discuss+unsubscribe@tensorflow.org.

Martin Wicke

unread,

Jul 6, 2018, 11:04:59 AM7/6/18

to Peng Yu, Discuss

And remove the "with graph" statement.

To unsubscribe from this group and stop receiving emails from it, send an email to discuss+u...@tensorflow.org.

Peng Yu

unread,

Jul 6, 2018, 1:34:39 PM7/6/18

to Martin Wicke, Discuss

In [3]: with tf.Session() as sess:

...: x = tf.placeholder(tf.int32, [None])

...: a = x + 1

...: y = tf.placeholder(tf.int32, [None])

...: b = y + 1

...: print(sess.run(b, feed_dict={y:a, x:[1]}))

...:

2018-07-06 13:34:00.518855: I tensorflow/core/platform/cpu_feature_guard.cc:137] Your CPU supports instructions that this TensorFlow binary was not compiled to use: SSE4.2 AVX AVX2 FMA

---------------------------------------------------------------------------

TypeError Traceback (most recent call last)

<ipython-input-3-47eee1c199ef> in <module>()

4 y = tf.placeholder(tf.int32, [None])

5 b = y + 1

----> 6 print(sess.run(b, feed_dict={y:a, x:[1]}))

7

~/.pyenv/virtualenvs/image-vectorization/3.6.1/lib/python3.6/site-packages/tensorflow/python/client/session.py in run(self, fetches, feed_dict, options, run_metadata)

893 try:

894 result = self._run(None, fetches, feed_dict, options_ptr,

--> 895 run_metadata_ptr)

896 if run_metadata:

897 proto_data = tf_session.TF_GetBuffer(run_metadata_ptr)

~/.pyenv/virtualenvs/image-vectorization/3.6.1/lib/python3.6/site-packages/tensorflow/python/client/session.py in _run(self, handle, fetches, feed_dict, options, run_metadata)

1074

1075 if isinstance(subfeed_val, ops.Tensor):

-> 1076 raise TypeError('The value of a feed cannot be a tf.Tensor object. '

1077 'Acceptable feed values include Python scalars, '

1078 'strings, lists, numpy ndarrays, or TensorHandles.')

TypeError: The value of a feed cannot be a tf.Tensor object. Acceptable feed values include Python scalars, strings, lists, numpy ndarrays, or TensorHandles.

To unsubscribe from this group and stop receiving emails from it, send an email to discuss+unsubscribe@tensorflow.org.

Martin Wicke

unread,

Jul 6, 2018, 1:35:41 PM7/6/18

to Peng Yu, Derek Murray, Discuss

+Derek Murray Now I'm confused.

To unsubscribe from this group and stop receiving emails from it, send an email to discuss+u...@tensorflow.org.

Reply all

Reply to author

Forward