Am I correct that in Tensorflow, when I run anything, my feed_dict needs to give values to all my placeholders, even ones that are irrelevant to what I'm running?
In particular I'm thinking of making a prediction, in which case my targets placeholder is irrelevant.
Well, it depends on how your computation graph looks like and how you run the ops which are fed by tensors (here: placeholders). If there's no dependency on the placeholder in any part of the computation graph that you'll execute in the session, then it does not need to be fed a value. Here's a small example:
In [90]: a = tf.constant([5, 5, 5], tf.float32, name='A')
...: b = tf.placeholder(tf.float32, shape=[3], name='B')
...: c = tf.constant([3, 3, 3], tf.float32, name='C')
...: d = tf.add(a, c, name="Add")
...:
...: with tf.Session() as sess:
...: print(sess.run(d))
...:
# result
[8. 8. 8.]
On the other hand, if you execute a part of the computation graph which has a dependency on the placeholder then a value it must be fed else it will raise InvalidArgumentError. Here's an example demonstrating this:
In [89]: a = tf.constant([5, 5, 5], tf.float32, name='A')
...: b = tf.placeholder(tf.float32, shape=[3], name='B')
...: c = tf.add(a, b, name="Add")
...:
...: with tf.Session() as sess:
...: print(sess.run(c))
...:
Executing the above code, throws the following InvalidArgumentError
InvalidArgumentError: You must feed a value for placeholder tensor 'B' with dtype float and shape [3]
[[Node: B = Placeholderdtype=DT_FLOAT, shape=[3], _device="/job:localhost/replica:0/task:0/device: CPU:0"]]
So, to make it work, you've to feed the placeholder using feed_dict as in:
In [91]: a = tf.constant([5, 5, 5], tf.float32, name='A')
...: b = tf.placeholder(tf.float32, shape=[3], name='B')
...: c = tf.add(a, b, name="Add")
...:
...: with tf.Session() as sess:
...: print(sess.run(c, feed_dict={b: [3, 3, 3]}))
...:
...:
[8. 8. 8.]