Tensorflow partial_run() “Must run 'setup' before performing partial runs!” despite being set up

Question

I am building a proof-of-concept around running sub-graphs without recomputing, using tensorflow\'s partial_run() methods.

Currently I have a simple little python script

Answer 1

The error message is quite explicit: You must run the setup before (every) partial run. I think this is the reference.
Additionally your ilist list of dicts only works for 2 runs. In the third run you only feed a value for a - this won't work. Here is a sample loop that works for me:

ilist = [{a: 1, b: 1}, {a: 2, b: 2}, {a: 1, b: 1}, {a: 1, b: 3}]

with tf.Session() as sess:
    for i, fd in enumerate(ilist):
        hdle = sess.partial_run_setup([y], [a, b])
        y_r = sess.partial_run(hdle, y, feed_dict=fd)

        eout = fd[a] * fd[b] + 1
        print("got {}, expected {}".format(y_r, eout))

Answer 2

This example from API documentation works :

import tensorflow as tf
a = tf.placeholder(tf.float32, shape=[])
b = tf.placeholder(tf.float32, shape=[])
c = tf.placeholder(tf.float32, shape=[])
r1 = tf.add(a, b)
r2 = tf.multiply(r1, c)

with tf.Session() as sess:
    h = sess.partial_run_setup([r1, r2], [a, b, c])
    res = sess.partial_run(h, r1, feed_dict={a: 1, b: 2})
    res = sess.partial_run(h, r2, feed_dict={c: 2})        
    print(res) #prints 6.0

But if we add on more invocation it doesn't . If this doesn't work what is the points in using partial_run.

import tensorflow as tf
a = tf.placeholder(tf.float32, shape=[])
b = tf.placeholder(tf.float32, shape=[])
c = tf.placeholder(tf.float32, shape=[])
r1 = tf.add(a, b)
r2 = tf.multiply(r1, c)

with tf.Session() as sess:
    h = sess.partial_run_setup([r1, r2], [a, b, c])
    res = sess.partial_run(h, r1, feed_dict={a: 1, b: 2})
    res = sess.partial_run(h, r2, feed_dict={c: 2})
    res = sess.partial_run(h, r2, feed_dict={c: 3})

    print(res)
InvalidArgumentError: Must run 'setup' before performing partial runs!