如何在张量流中迭代张量？

Question

I am getting the following error:

TypeError: 'Tensor' object is not iterable.

I am trying to use a placeholder and FIFOQueue to feed data. But the problem here is I cannot batch the data. Could any one provide a solution?

I'm new in TensorFlow and confused the concept of placeholder and tensor.

Here is the code:

#-*- coding:utf-8 -*-  
import tensorflow as tf  
import sys  

q = tf.FIFOQueue(1000,tf.string)  
label_ph = tf.placeholder(tf.string,name="label")
enqueue_op = q.enqueue_many(label_ph)
qr = tf.train.QueueRunner(q,enqueue_op)  
m = q.dequeue()

sess_conf = tf.ConfigProto()
sess_conf.gpu_options.allow_growth = True
sess = tf.Session(config=sess_conf)  
sess.run(tf.global_variables_initializer())  
coord = tf.train.Coordinator()  
tf.train.start_queue_runners(coord=coord, sess=sess) 

image_batch = tf.train.batch(
        m,batch_size=3,
        enqueue_many=True,
        capacity=9
        )

for i in range(0, 10):  
    print "-------------------------"  
    #print(sess.run(q.dequeue()))  
    a = ['a','b','c','a1','b1','c1','a','b','c2','a','b','c3',]
    sess.run(enqueue_op,{label_ph:a})
    b = sess.run(m)
    print b
q.close()
coord.request_stop()  

Answer 1

I think you're running into the same issue I am. When you run the session you don't actually have access to the data, you have access to the data graph instead. So you should think of the tensor objects like nodes in a graph and not like chunks of data you can do things with. If you want to do things to the graph you have to call tf.* functions or get the variables inside a sess.run() call. When you do that then tensorflow will figure out how to get that data based on its dependencies and run the computation.

As for your question, check out the QueueRunner example on this page. https://www.tensorflow.org/programmers_guide/threading_and_queues

Another way you could do it (which is what I switched to) is that you could shuffle your data on the cpu and then copy it all over at once. Then you keep track of what step you're on and grab the data for that step. I helps to keep the gpu feed with data and reduces memory copies.

    all_shape = [num_batches, batch_size, data_len]
    local_shape = [batch_size, data_len]

    ## declare the data variables
    model.all_data = tf.Variable(tf.zeros(all_shape), dtype=tf.float32)
    model.step_data=tf.Variable(tf.zeros(local_shape), dtype=tf.float32)
    model.step = tf.Variable(0, dtype=tf.float32, trainable=False, name="step")

    ## then for each step in the epoch grab the data
    index = tf.to_int32(model.step)
    model.step_data = model.all_data[index]

    ## inc the step counter
    tf.assign_add(model.step, 1.0, name="inc_counter")