tensorflow: Why doesn't tensorflow support tensor as the feed_dict?

What I’m trying to do

I am trying to extract CNN features for my own images with residual-net based on https://github.com/ry/tensorflow-resnet. I plan to input image data from JPG files before exploring how to convert the images into a single file.

What I have done

I have read https://www.tensorflow.org/versions/r0.9/how_tos/reading_data/index.html and some related materials about how to input data like feeding and placeholder. Here is my code:

import tensorflow as tf
from convert import print_prob, checkpoint_fn, meta_fn
from image_processing import image_preprocessing
tf.app.flags.DEFINE_integer('batch_size', 1, "batch size")
tf.app.flags.DEFINE_integer('input_size', 224, "input image size")
tf.app.flags.DEFINE_integer('min_after_dequeue', 224, "min after dequeue")
tf.app.flags.DEFINE_integer('layers', 152, "The number of layers in the net")
tf.app.flags.DEFINE_integer('image_number', 6951, "number of images")
FLAGS = tf.app.flags.FLAGS


def placeholder_inputs():
    images_placeholder = tf.placeholder(tf.float32, shape=(FLAGS.batch_size, FLAGS.input_size, FLAGS.input_size, 3))
    label_placeholder = tf.placeholder(tf.int32, shape=FLAGS.batch_size)
    return images_placeholder, label_placeholder


def fill_feed_dict(image_ba, label_ba, images_pl, labels_pl):
    feed_dict = {
        images_pl: image_ba,
    }
    return feed_dict

min_fraction_of_examples_in_queue = 0.4
min_queue_examples = int(FLAGS.image_number *
                     min_fraction_of_examples_in_queue)
dataset = tf.train.string_input_producer(["hollywood_test.txt"])
reader = tf.TextLineReader()
_, file_content = reader.read(dataset)
image_name, label, _ = tf.decode_csv(file_content, [[""], [""], [""]], " ")
label = tf.string_to_number(label)
num_preprocess_threads = 10
images_and_labels = []
with tf.Session() as sess:
    for thread_id in range(num_preprocess_threads):
        image_buffer = tf.read_file(image_name)
        bbox = []
        train = False
        image = image_preprocessing(image_buffer, bbox, train, thread_id)
        image = image_buffer
        images_and_labels.append([image, label])
    image_batch, label_batch = tf.train.batch_join(images_and_labels,
                                            batch_size=FLAGS.batch_size,
                                            capacity=min_queue_examples + 3 * FLAGS.batch_size)
    images_placeholder, labels_placeholder = placeholder_inputs()
    new_saver = tf.train.import_meta_graph(meta_fn(FLAGS.layers))
    new_saver.restore(sess, checkpoint_fn(FLAGS.layers))
    graph = tf.get_default_graph()
    prob_tensor = graph.get_tensor_by_name("prob:0")
    images = graph.get_tensor_by_name("images:0")
    feed_dict = fill_feed_dict(image_batch, label_batch, images, labels_placeholder)
    coord = tf.train.Coordinator()
    threads = tf.train.start_queue_runners(coord=coord)
    sess.run(tf.initialize_all_variables())
    prob = sess.run(prob_tensor, feed_dict=feed_dict)
    print_prob(prob[0])
    coord.request_stop()
    coord.join(threads)

What my question is

The code above got the error TypeError: The value of a feed cannot be a tf.Tensor object. Acceptable feed values include Python scalars, strings, lists, or numpy ndarrays. You can see I am trying to do all the work in the context of tensorflow. As far as I know, tensorflow is a framework where all the inputs and outputs of the nodes are tensors. However, I am quite confused why feed_dict doesn’t support tensor as an input. But the batch_join function return a tensor and so do other ops. I found that in the mnist example of tensorflow there is even another function to produce a batch when tensorflow is providing methods for batching. So I wonder if there is an elegant way to do these things. If this is because of my lack of searching and careful reading I really apologize.