tensorflow: Reading data with queue is even slower than using feed_dict

It seems reading data with feed_dict is inefficient in tensorflow (I found it 3-4 times slower than a theano-based implementation with a totally same network structure), I turn to a queue-based method as official recommendation. However, my experiment result give a even worse performance. Is there anything wrong ?

Related issue

#3377 Moving data from CPU to GPU is slow

Environment info

Operating System: Ubuntu 16.04 CPU: Intel i4790-k GPU: Nvidia GTX 1070 (8 GB) Memory: 16 GB Tensorflow version: 1.0 CUDA version: 8.0 cuDNN version: 5.0

Implementation Example

1. Using feed_dict

x, y = np.array(...)
in_x, in_y = tf.placeholder(...)
""" f() is some computation in the network, including embedding_lookup, 
bidirectional_dynamic_rnn, dense """
train_op = f(in_x, in_y) 
sess = tf.Session()
for _ in range(num_epochs): # num_epochs is 100 here
    sess.run(train_op, {in_x: x, in_y: y})

1. Using queue

x, y = np.array(...)
x, y = tf.convert_to_tensor(...)
in_x, in_y = tf.train.slice_input_producer([x, y], num_epochs=100)
in_x, in_y = tf.train.batch([in_x, in_y], batch_size=32)
train_op = f(in_x, in_y)
sess = tf.Session()
coord = tf.train.Coordinator()
threads = tf.train.start_queue_runners(sess, coord)
try:
    while not coord.should_stop():
        sess.run(train_op)
    except Exception as e:
        coord.request_stop(e)
    finally:
        coord.request_stop()
coord.join(threads)