Issue feeding a list into feed_dict in TensorFlow

feed_dict can be provided by preparing a dictionary beforehand as follows

n = 10
input_1 = [tf.placeholder(...) for _ in range(n)]
input_2 = tf.placeholder(...)
data_1 = [np.array(...) for _ in range(n)]
data_2 = np.array(...)


feed_dictionary = {}
for i in range(n):
    feed_dictionary[input_1[i]] = data_1[i]
feed_dictionary[input_2] = data_2
sess.run(y, feed_dict=feed_dictionary)

Here is a correct example:

batch_size, input_size, n = 2, 3, 2
# in your case n = 10
x = tf.placeholder(tf.types.float32, shape=(n, batch_size, input_size))
y = tf.add(x, x)

data = np.random.rand(n, batch_size, input_size)

sess = tf.Session()
print sess.run(y, feed_dict={x: data})

And here is a strange things I see in your approach. For some reason you use 10 * [tf.placeholder(...)], which creates 10 tensors of size (batch_size, input_size). No idea why do you do this, if you can just create on Tensor of rank 3 (where the first dimension is 10).

Because you have a list of tensors (and not a tensor), you can not feed your data to this list (but in my case I can feed to my tensor).


There are two issues that are causing problems here:

The first issue is that the Session.run() call only accepts a small number of types as the keys of the feed_dict. In particular, lists of tensors are not supported as keys, so you have to put each tensor as a separate key.* One convenient way to do this is using a dictionary comprehension:

inputs = [tf.placeholder(...), ...]
data = [np.array(...), ...]
sess.run(y, feed_dict={i: d for i, d in zip(inputs, data)})

The second issue is that the 10 * [tf.placeholder(...)] syntax in Python creates a list with ten elements, where each element is the same tensor object (i.e. has the same name property, the same id property, and is reference-identical if you compare two elements from the list using inputs[i] is inputs[j]). This explains why, when you tried to create a dictionary using the list elements as keys, you ended up with a dictionary with a single element - because all of the list elements were identical.

To create 10 different placeholder tensors, as you intended, you should instead do the following:

inputs = [tf.placeholder(tf.float32, shape=(batch_size, input_size))
          for _ in xrange(10)]

If you print the elements of this list, you'll see that each element is a tensor with a different name.


EDIT: * You can now pass tuples as the keys of a feed_dict, because these may be used as dictionary keys.