Tensorflow TypeError: Fetch argument None has inva

Question

I'm building a RNN loosely based on the TensorFlow tutorial.

The relevant parts of my model are as follows:

input_sequence = tf.placeholder(tf.float32, [BATCH_SIZE, TIME_STEPS, PIXEL_COUNT + AUX_INPUTS])
output_actual = tf.placeholder(tf.float32, [BATCH_SIZE, OUTPUT_SIZE])

lstm_cell = tf.nn.rnn_cell.BasicLSTMCell(CELL_SIZE, state_is_tuple=False)
stacked_lstm = tf.nn.rnn_cell.MultiRNNCell([lstm_cell] * CELL_LAYERS, state_is_tuple=False)

initial_state = state = stacked_lstm.zero_state(BATCH_SIZE, tf.float32)
outputs = []

with tf.variable_scope("LSTM"):
    for step in xrange(TIME_STEPS):
        if step > 0:
            tf.get_variable_scope().reuse_variables()
        cell_output, state = stacked_lstm(input_sequence[:, step, :], state)
        outputs.append(cell_output)

final_state = state

And the feeding:

cross_entropy = tf.reduce_mean(-tf.reduce_sum(output_actual * tf.log(prediction), reduction_indices=[1]))
train_step = tf.train.AdamOptimizer(learning_rate=LEARNING_RATE).minimize(cross_entropy)
correct_prediction = tf.equal(tf.argmax(prediction, 1), tf.argmax(output_actual, 1))
accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))

with tf.Session() as sess:
    sess.run(tf.initialize_all_variables())
    numpy_state = initial_state.eval()

    for i in xrange(1, ITERATIONS):
        batch = DI.next_batch()

        print i, type(batch[0]), np.array(batch[1]).shape, numpy_state.shape

        if i % LOG_STEP == 0:
            train_accuracy = accuracy.eval(feed_dict={
                initial_state: numpy_state,
                input_sequence: batch[0],
                output_actual: batch[1]
            })

            print "Iteration " + str(i) + " Training Accuracy " + str(train_accuracy)

        numpy_state, train_step = sess.run([final_state, train_step], feed_dict={
            initial_state: numpy_state,
            input_sequence: batch[0],
            output_actual: batch[1]
            })

When I run this, I get the following error:

Traceback (most recent call last):
  File "/home/agupta/Documents/Projects/Image-Recognition-with-LSTM/RNN/feature_tracking/model.py", line 109, in <module>
    output_actual: batch[1]
  File "/usr/local/lib/python2.7/dist-packages/tensorflow/python/client/session.py", line 698, in run
    run_metadata_ptr)
  File "/usr/local/lib/python2.7/dist-packages/tensorflow/python/client/session.py", line 838, in _run
    fetch_handler = _FetchHandler(self._graph, fetches)
  File "/usr/local/lib/python2.7/dist-packages/tensorflow/python/client/session.py", line 355, in __init__
    self._fetch_mapper = _FetchMapper.for_fetch(fetches)
  File "/usr/local/lib/python2.7/dist-packages/tensorflow/python/client/session.py", line 181, in for_fetch
    return _ListFetchMapper(fetch)
  File "/usr/local/lib/python2.7/dist-packages/tensorflow/python/client/session.py", line 288, in __init__
    self._mappers = [_FetchMapper.for_fetch(fetch) for fetch in fetches]
  File "/usr/local/lib/python2.7/dist-packages/tensorflow/python/client/session.py", line 178, in for_fetch
    (fetch, type(fetch)))
TypeError: Fetch argument None has invalid type <type 'NoneType'>

Perhaps the weirdest part is that this error gets thrown the second iteration, and the first works completely fine. I'm ripping my hair trying to fix this, so any help would be greatly appreciated.

Answer 1

You are re-assigning the train_step variable to the second element of the result of sess.run() (which happens to be None). Hence, on the second iteration, train_step is None, which leads to the error.

The fix is fortunately simple:

for i in xrange(1, ITERATIONS):

    # ...

    # Discard the second element of the result.
    numpy_state, _ = sess.run([final_state, train_step], feed_dict={
        initial_state: numpy_state,
        input_sequence: batch[0],
        output_actual: batch[1]
        })

Answer 2

Another common reason to get this error is if you include the summary fetch operation but have not written any summaries.

Example:

# tf.summary.scalar("loss", loss) # <- uncomment this line and it will work fine
summary_op = tf.summary.merge_all()
sess = tf.Session()
# ...
summary = sess.run([summary_op, ...], feed_dict={...}) # TypeError, summary_op is "None"!

What's extra confusing is that summary_op is not itself None, that's just the error that bubbles up from inside the session's run method.