21

I am trying to implement a simple feed forward network. However, I can't figure out how to feed a Placeholder. This example:

import tensorflow as tf

num_input  = 2
num_hidden = 3
num_output = 2

x  = tf.placeholder("float", [num_input, 1])
W_hidden = tf.Variable(tf.zeros([num_hidden, num_input]))
W_out    = tf.Variable(tf.zeros([num_output, num_hidden]))
b_hidden = tf.Variable(tf.zeros([num_hidden]))
b_out    = tf.Variable(tf.zeros([num_output]))

h = tf.nn.softmax(tf.matmul(W_hidden,x) + b_hidden)

sess = tf.Session()

with sess.as_default():
    print h.eval()

Gives me the following error:

  ...
    results = self._do_run(target_list, unique_fetch_targets, feed_dict_string)
  File "/usr/local/lib/python2.7/dist-packages/tensorflow/python/client/session.py", line 419, in _do_run
    e.code)
tensorflow.python.framework.errors.InvalidArgumentError: You must feed a value for placeholder tensor 'Placeholder' with dtype float and shape dim { size: 2 } dim { size: 1 }
     [[Node: Placeholder = Placeholder[dtype=DT_FLOAT, shape=[2,1], _device="/job:localhost/replica:0/task:0/cpu:0"]()]]
Caused by op u'Placeholder', defined at:
  File "/home/sfalk/workspace/SemEval2016/java/semeval2016-python/slot1_tf.py", line 8, in <module>
    x  = tf.placeholder("float", [num_input, 1])
  ...

I have tried

tf.assign([tf.Variable(1.0), tf.Variable(1.0)], x)
tf.assign([1.0, 1.0], x)

but that does not work apparently.

Stefan Falk
  • 23,898
  • 50
  • 191
  • 378

1 Answers1

34

To feed a placeholder, you use the feed_dict argument to Session.run() (or Tensor.eval()). Let's say you have the following graph, with a placeholder:

x = tf.placeholder(tf.float32, shape=[2, 2])
y = tf.constant([[1.0, 1.0], [0.0, 1.0]])
z = tf.matmul(x, y)

If you want to evaluate z, you must feed a value for x. You can do this as follows:

sess = tf.Session()
print sess.run(z, feed_dict={x: [[3.0, 4.0], [5.0, 6.0]]})

For more information, see the documentation on feeding.

mrry
  • 125,488
  • 26
  • 399
  • 400
  • Hmm.. is there no other way? That looks inconvenient if I e.g. want to look at intermediary results. – Stefan Falk Nov 19 '15 at 19:56
  • You can also pass the `feed_dict` argument to `Tensor.eval()`, which might be more convenient when building the graph. If you want a "sticky" placeholder, I'd suggest making your own function that wraps `sess.run()`, captures a set of feed values, and passes that to the `run()` call each time. – mrry Nov 19 '15 at 20:15
  • @mrry, could you make an example of your comment? Thanks – Amir Aug 23 '17 at 03:27