Question

我有一个图像的名称和标签作为列表，我想得到一批64个图像/标签。我可以以正确的方式获得图像，但对于标签，其尺寸为（64,8126）。每列具有64次相同的元素。并且行包含8126个原始标签值而不会被洗牌。

我理解每个图像tf.train.shuffle_batch都考虑8126元素标签向量的问题。但是，我如何只为每个图像传递单个元素？

def _get_images(shuffle=True):

"""Gets the images and labels as a batch"""

    #get image and label list
    _img_names,_img_class = _get_list() #list of image names and labels

    filename_queue = tf.train.string_input_producer(_img_names)

    #reader
    image_reader = tf.WholeFileReader()
    _, image_file = image_reader.read(filename_queue)

    #decode jpeg
    image_original = tf.image.decode_jpeg(image_file)
    label_original = tf.convert_to_tensor(_img_class,dtype=tf.int32)
    #print label_original

    #image preprocessing
    image = tf.image.resize_images(image_original, [224,224])
    float_image = tf.cast(image,dtype=tf.float32)
    float_image = tf.image.per_image_standardization(image)
    #set the shape
    float_image.set_shape((224, 224, 3))
    #label_original.set_shape([8126]) #<<<<<=========== causes (64,8126) dimension label without shuffle

    #parameters for shuffle
    batch_size = 64
    num_preprocess_threads = 16
    num_examples_per_epoch = 8000
    min_fraction_of_examples_in_queue = 0.4
    min_queue_examples = int(num_examples_per_epoch *
                       min_fraction_of_examples_in_queue)

    if shuffle:
        images_batch, label_batch = tf.train.shuffle_batch(
            [float_image,label_original],
            batch_size=batch_size,
            num_threads=num_preprocess_threads,
            capacity=min_queue_examples + 3 * batch_size,
            min_after_dequeue=min_queue_examples)
    else:
        images_batch, label_original = tf.train.batch(
            [float_image,_img_class],
            batch_size=batch_size,
            num_threads=num_preprocess_threads,
            capacity=min_queue_examples + 3 * batch_size)

    return images_batch,label_batch

Answer 1

您可以使用tf.train.slice_input_producer

# here _img_class should be a list
labels_queue = tf.train.slice_input_producer([_img_class])
...
images_batch, label_batch = tf.train.shuffle_batch(
        [float_image,labels_queue],
        batch_size=batch_size,
        num_threads=num_preprocess_threads,
        capacity=min_queue_examples + 3 * batch_size,
        min_after_dequeue=min_queue_examples)

如何处理输入管道中的标签？

1 个答案: