5

目的是创建一个 TfRecords 数据库。给定：我有 23 个文件夹，每个文件夹包含 7500 个图像和 23 个文本文件，每个文件都有 7500 行描述不同文件夹中 7500 个图像的特征。

我通过以下代码创建了数据库：

import tensorflow as tf
import numpy as np
from PIL import Image

def _Float_feature(value):
    return tf.train.Feature(float_list=tf.train.FloatList(value=[value]))

def _bytes_feature(value):
    return tf.train.Feature(bytes_list=tf.train.BytesList(value=[value]))

def _int64_feature(value):
    return tf.train.Feature(int64_list=tf.train.Int64List(value=[value]))

def create_image_annotation_data():
    # Code to read images and features.
    # images represent a list of numpy array of images, and features_labels represent a list of strings
    # where each string represent the whole set of features for each image. 
    return images, features_labels

# This is the starting point of the program.
# Now I have the images stored as list of numpy array, and the features as list of strings.
images, annotations = create_image_annotation_data()

tfrecords_filename = "database.tfrecords"
writer = tf.python_io.TFRecordWriter(tfrecords_filename)

for img, ann in zip(images, annotations):

    # Note that the height and width are needed to reconstruct the original image.
    height = img.shape[0]
    width = img.shape[1]

    # This is how data is converted into binary
    img_raw = img.tostring()
    example = tf.train.Example(features=tf.train.Features(feature={
        'height': _int64_feature(height),
        'width': _int64_feature(width),
        'image_raw': _bytes_feature(img_raw),
        'annotation_raw': _bytes_feature(tf.compat.as_bytes(ann))
    }))

    writer.write(example.SerializeToString())

writer.close()

reconstructed_images = []

record_iterator = tf.python_io.tf_record_iterator(path=tfrecords_filename)

for string_record in record_iterator:
    example = tf.train.Example()
    example.ParseFromString(string_record)

    height = int(example.features.feature['height']
                 .int64_list
                 .value[0])

    width = int(example.features.feature['width']
                .int64_list
                .value[0])

    img_string = (example.features.feature['image_raw']
                  .bytes_list
                  .value[0])

    annotation_string = (example.features.feature['annotation_raw']
                         .bytes_list
                         .value[0])

    img_1d = np.fromstring(img_string, dtype=np.uint8)
    reconstructed_img = img_1d.reshape((height, width, -1))
    annotation_reconstructed = annotation_string.decode('utf-8')

因此，在将图像和文本转换为 tfRecords 并能够读取它们并将图像转换为 numpy 并将（二进制文本）转换为 python 中的字符串之后，我尝试通过使用带有阅读器的 filename_queue 来加倍努力（目的是为图提供一批数据，而不是一次提供一组数据。此外，目的是通过不同的线程对示例队列进行入队和出队，因此，可以更快地训练网络）

因此，我使用了以下代码：

import tensorflow as tf
import numpy as np
import time

image_file_list = ["database.tfrecords"]
batch_size = 16

# Make a queue of file names including all the JPEG images files in the relative
# image directory.
filename_queue = tf.train.string_input_producer(image_file_list, num_epochs=1, shuffle=False)

reader = tf.TFRecordReader()

# Read a whole file from the queue, the first returned value in the tuple is the
# filename which we are ignoring.
_, serialized_example = reader.read(filename_queue)

features = tf.parse_single_example(
      serialized_example,
      # Defaults are not specified since both keys are required.
      features={
          'height': tf.FixedLenFeature([], tf.int64),
          'width': tf.FixedLenFeature([], tf.int64),
          'image_raw': tf.FixedLenFeature([], tf.string),
          'annotation_raw': tf.FixedLenFeature([], tf.string)
      })

image = tf.decode_raw(features['image_raw'], tf.uint8)
annotation = tf.decode_raw(features['annotation_raw'], tf.float32)

height = tf.cast(features['height'], tf.int32)
width = tf.cast(features['width'], tf.int32)

image = tf.reshape(image, [height, width, 3])

# Note that the minimum after dequeue is needed to make sure that the queue is not empty after dequeuing so that
# we don't run into errors
'''
min_after_dequeue = 100
capacity = min_after_dequeue + 3 * batch_size
ann, images_batch = tf.train.batch([annotation, image],
                                   shapes=[[1], [112, 112, 3]],
                                   batch_size=batch_size,
                                   capacity=capacity,
                                   num_threads=1)
'''

# Start a new session to show example output.
with tf.Session() as sess:
    merged = tf.summary.merge_all()
    train_writer = tf.summary.FileWriter('C:/Users/user/Documents/tensorboard_logs/New_Runs', sess.graph)

    # Required to get the filename matching to run.
    tf.global_variables_initializer().run()

    # Coordinate the loading of image files.
    coord = tf.train.Coordinator()
    threads = tf.train.start_queue_runners(coord=coord)

    for steps in range(16):
        t1 = time.time()
        annotation_string, batch, summary = sess.run([annotation, image, merged])
        t2 = time.time()
        print('time to fetch 16 faces:', (t2 - t1))
        print(annotation_string)
        tf.summary.image("image_batch", image)
        train_writer.add_summary(summary, steps)

    # Finish off the filename queue coordinator.
    coord.request_stop()
    coord.join(threads)

最后，运行上述代码后，出现如下错误： OutOfRangeError（回溯见上文）：FIFOQueue '_0_input_producer' is closed and has enough elements (requested 1, current size 0) [[Node: ReaderReadV2 = ReaderReadV2[_device= "/job:localhost/replica:0/task:0/cpu:0"](TFRecordReaderV2, input_producer)]]

另一个问题：

  1. 如何解码二进制数据库（tfrecords）以检索“作为python字符串数据结构”存储的特征。
  2. 如何使用tf.train.batch创建一批示例来馈送网络。

谢谢！！任何帮助深表感谢。

4

1 回答 1

4

为了解决这个问题，coordinator必须queue runner在Session. 此外，由于 epoch 的数量是内部控制的，所以它不是 a global variable，而是考虑 a local variable。因此，我们需要先初始化该局部变量，然后再告诉queue_runner开始将file_names入队Queue。因此，这里是以下代码：

filename_queue = tf.train.string_input_producer(tfrecords_filename, num_epochs=num_epoch, shuffle=False, name='queue')
reader = tf.TFRecordReader()

key, serialized_example = reader.read(filename_queue)
features = tf.parse_single_example(
    serialized_example,
    # Defaults are not specified since both keys are required.
    features={
        'height': tf.FixedLenFeature([], tf.int64),
        'width': tf.FixedLenFeature([], tf.int64),
        'image_raw': tf.FixedLenFeature([], tf.string),
        'annotation_raw': tf.FixedLenFeature([], tf.string)
    })
...
init_op = tf.group(tf.local_variables_initializer(),
               tf.global_variables_initializer())
with tf.Session() as sess:
    sess.run(init_op)

    coord = tf.train.Coordinator()
    threads = tf.train.start_queue_runners(coord=coord)

现在应该可以了。

现在要在将它们输入网络之前收集一批图像，我们可以使用tf.train.shuffle_batchor tf.train.batch。两者都有效。区别很简单。一个洗牌图像，另一个不洗牌。但请注意，定义一个线程并使用线程tf.train.batch可能会打乱数据样本，因为正在排队的线程之间存在竞争file_names。无论如何，应该在初始化之后直接插入以下代码Queue，如下所示：

min_after_dequeue = 100
num_threads = 1
capacity = min_after_dequeue + num_threads * batch_size
label_batch, images_batch = tf.train.batch([annotation, image],
                                       shapes=[[], [112, 112, 3]],
                                       batch_size=batch_size,
                                       capacity=capacity,
                                       num_threads=num_threads)

请注意，这里的形状tensors可能会有所不同。碰巧读者正在解码一个大小为 的彩色图像[112, 112, 3]。并且注释有一个[]（没有理由，那是一个特例）。

最后，我们可以将tf.string数据类型视为字符串。实际上，在评估标注张量之后，我们可以意识到张量被视为一个binary string（这就是它在 tensorflow 中的真正处理方式）。因此，就我而言，该字符串只是与该特定图像相关的一组特征。因此，为了提取特定的特征，这里有一个例子：

# The output of string_split is not a tensor, instead, it is a SparseTensorValue. Therefore, it has a property value that stores the actual values. as a tensor. 
label_batch_splitted = tf.string_split(label_batch, delimiter=', ')
label_batch_values = tf.reshape(label_batch_splitted.values, [batch_size, -1])
# string_to_number will convert the feature's numbers into float32 as I need them. 
label_batch_numbers = tf.string_to_number(label_batch_values, out_type=tf.float32)
# the tf.slice would extract the necessary feature which I am looking.
confidences = tf.slice(label_batch_numbers, begin=[0, 3], size=[-1, 1])

希望这个答案有帮助。

于 2017-05-21T17:48:23.823 回答