KevinXu tutorial-TFRecord-notMINST datasets

这部分主要是讲怎么把自己的数据转换成TFRecord

主要步骤如下：

1.转换数据格式

2.读取数据以及解码

3.生成Batch

4.构建tensorflow图谱

5.训练模型/验证、测试

原作者Kevin的视频课程地址：Youtube

关于TFRecord

是tensorflow的官方数据格式，请看截图，不是这部分的重点，略过

关于dataset：notMINST

下载数据集

数据集地址：点击打开链接

有以上两种可供选择，large的数据集比较大，包含所有数据。small每个类别有1873张，看自己选择吧，我用small。

--------------------------------------代码分解-----------------------------------------------

1.导入库

import tensorflow as tf
import numpy as np
import os
import matplotlib.pyplot as plt
import skimage.io as io   # anaconda 自带

2.通过数据文件地址得到images and labels

def get_file(file_dir):
    '''Get full image directory and corresponding labels
    Args:
        file_dir: file directory
    Returns:
        images: image directories, list, string
        labels: label, list, int
    '''

    images = []
    temp = []
    for root, sub_folders, files in os.walk(file_dir):
        # image directories
        for name in files:
            images.append(os.path.join(root, name))
        # get 10 sub-folder names
        for name in sub_folders:
            temp.append(os.path.join(root, name))

    # assign 10 labels based on the folder names
    labels = []
    for one_folder in temp:
        n_img = len(os.listdir(one_folder))
        letter = one_folder.split('/')[-1]
        # 通过子文件夹名字获取标签，标签编号为1-10

        if letter == 'A':
            labels = np.append(labels, n_img * [1])
        elif letter == 'B':
            labels = np.append(labels, n_img * [2])
        elif letter == 'C':
            labels = np.append(labels, n_img * [3])
        elif letter == 'D':
            labels = np.append(labels, n_img * [4])
        elif letter == 'E':
            labels = np.append(labels, n_img * [5])
        elif letter == 'F':
            labels = np.append(labels, n_img * [6])
        elif letter == 'G':
            labels = np.append(labels, n_img * [7])
        elif letter == 'H':
            labels = np.append(labels, n_img * [8])
        elif letter == 'I':
            labels = np.append(labels, n_img * [9])
        else:
            labels = np.append(labels, n_img * [10])

    # shuffle
    temp = np.array([images, labels])
    temp = temp.transpose()
    np.random.shuffle(temp)

    image_list = list(temp[:, 0])
    label_list = list(temp[:, 1])
    label_list = [int(float(i)) for i in label_list]

    return image_list, label_list

3.将标签和图像转为特定的格式：
因为后面convert_to_tfrecord函数用到的编码函数tf.train.Example的数据结构中包含了一个从属性到取值的字典。而属性的取值可以为：字符串列表(BytesList)、实数列表(FloatList)或者整数列表(Int64List)通过以下函数编码为Example proto形式的返回值，所以提前在这里转换。

# 以下两个函数来自tensorflow官网

def int64_feature(value):
  """Wrapper for inserting int64 features into Example proto."""
  if not isinstance(value, list):
    value = [value]
  return tf.train.Feature(int64_list=tf.train.Int64List(value=value))

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

4.将数据转换成TFRecord

def convert_to_tfrecord(images, labels, save_dir, name):
    '''convert all images and labels to one tfrecord file.
    Args:
        images: list of image directories, string type
        labels: list of labels, int type
        save_dir: the directory to save tfrecord file, e.g.: '/home/folder1/'
        name: the name of tfrecord file, string type, e.g.: 'train'
    Return:
        no return
    Note:
        converting needs some time, be patient...
        转换时间有点长
    '''

    # 指定数据转换格式后的保存路径和名称
    filename = os.path.join(save_dir, name + '.tfrecords')
    # 取得图片的样本总数
    n_samples = len(labels)

    # 检查图像和label是不是对应的
    if np.shape(images)[0] != n_samples:
        raise ValueError('Images size %d does not match label size %d.' % (images.shape[0], n_samples))

    # wait some time here, transforming need some time based on the size of your data.
    # 创建一个实例对象 writer，用于后面序列化数据的写入
    writer = tf.python_io.TFRecordWriter(filename)
    print('\nTransform start......')
    # 将所有数据(包括标签等)按照 tf.train.Example Protocol Buffer 的格式存储
    for i in np.arange(0, n_samples):
        try:
            image = io.imread(images[i])  # type(image) must be array! 一张一张的读取
            image_raw = image.tostring()   # 将图片矩阵转化为字符串
            label = int(labels[i])

            # 创建tf.train.Example 协议内存块，把标签、图片数据作为特定字段存入（数据类型转换）
            example = tf.train.Example(features=tf.train.Features(feature={
                'label': int64_feature(label),
                'image_raw': bytes_feature(image_raw)}))

            # 调用实例对象 writer 的 write 方法将序列化后的 example 协议内存块写入 TFRecord 文件
            writer.write(example.SerializeToString())
        except IOError as e:
            print('Could not read:', images[i])
            print('error: %s' % e)
            print('Skip it!\n')
            # 原始数据有几张图片有错误不能读取，用这种方式略过
    # 调用实例对象 writer 的 close 方法结束写入过程
    writer.close()
    print('Transform done!')

5. 读取 、解码生成batch

# 以上将数据转换成了TFRecord形式，将其解码出来
def read_and_decode(tfrecords_file, batch_size):
    '''read and decode tfrecord file, generate (image, label) batches
    Args:
        tfrecords_file: the directory of tfrecord file
        batch_size: number of images in each batch
    Returns:
        image: 4D tensor - [batch_size, width, height, channel]
        label: 1D tensor - [batch_size]
    '''
    # make an input queue from the tfrecord file 创建文件名队列，方便数据的读取并节约时间
    # 关于队列的原理见知乎地址： https://zhuanlan.zhihu.com/p/27238630
    filename_queue = tf.train.string_input_producer([tfrecords_file])

    # 创建一个实例对象 reader, 用于读取 TFRecord 中的样例
    reader = tf.TFRecordReader()
    _, serialized_example = reader.read(filename_queue)

    # 按照字段格式，解析数据，从TFRecord 到 tf.int64、tf.string
    img_features = tf.parse_single_example(
        serialized_example,
        features={
            'label': tf.FixedLenFeature([], tf.int64),
            'image_raw': tf.FixedLenFeature([], tf.string),
        })
    # 解码
    image = tf.decode_raw(img_features['image_raw'], tf.uint8)

    ##########################################################
    # you can put data augmentation here, I didn't use it
    ##########################################################
    # all the images of notMNIST are 28*28, you need to change the image size if you use other dataset.

    image = tf.reshape(image, [28, 28])
    label = tf.cast(img_features['label'], tf.int32)
    image_batch, label_batch = tf.train.batch([image, label],
                                              batch_size=batch_size,
                                              num_threads=64,
                                              capacity=2000)
    # num_threads：可以指定多个线程同时执行入队操作(数据读取和预处理)，通过队列实现多线程处理机制
    # capacity: 队列中最多可以存储的样例个数
    return image_batch, tf.reshape(label_batch, [batch_size])

6.以上就是数据的转换及读取，接下来测试以下刚才的代码

这部分注意自己的地址要更改过来，我的项目总览如下

# %% Convert data to TFRecord

test_dir = 'D://python//neural network//notMNIST//notMNIST_small//'
save_dir = 'D://python//neural network//notMNIST//'
BATCH_SIZE = 25

# Convert test data: you just need to run it ONCE !
name_test = 'test'
images, labels = get_file(test_dir)
convert_to_tfrecord(images, labels, save_dir, name_test)


# %% TO test train.tfrecord file

def plot_images(images, labels):
    '''plot one batch size
    '''
    for i in np.arange(0, BATCH_SIZE):
        plt.subplot(5, 5, i + 1)
        plt.axis('off')
        plt.title(chr(ord('A') + labels[i] - 1), fontsize=14)
        plt.subplots_adjust(top=1.5)
        plt.imshow(images[i])
    plt.show()


tfrecords_file = 'D://python//neural network//notMNIST//test.tfrecords'
# 这个地址为 save_file的地址加上数据的名字，详情见convert_to_tfrecord函数
image_batch, label_batch = read_and_decode(tfrecords_file, batch_size=BATCH_SIZE)

with tf.Session()  as sess:
    i = 0
    coord = tf.train.Coordinator()
    threads = tf.train.start_queue_runners(coord=coord)

    try:
        while not coord.should_stop() and i < 1:
            # just plot one batch size
            image, label = sess.run([image_batch, label_batch])
            plot_images(image, label)
            i += 1

    except tf.errors.OutOfRangeError:
        print('done!')
    finally:
        coord.request_stop()
    coord.join(threads)

7.报错可能：

tensorflow.python.framework.errors_impl.NotFoundError: Failed to create a NewWriteableFile: D:/python/neural network/notMNIST/test.tfrecords ϵͳ\udcd5Ҳ\udcbb\udcb5\udcbdָ\udcb6\udca8\udcb5\udcc4·\udcbe\udcb6\udca1\udca3

其实这是个傻瓜错误，因为路径给错了。把文件夹MINST写成了MNIST，以至于找不到文件，你们可要仔细哦，最好是复制路径，避免手写错误。

8.结果：转换后的image部分展示

附全部代码：

# 原作者相关信息
# By @Kevin Xu
# kevin28520@gmail.com
# My youtube: https://www.youtube.com/channel/UCVCSn4qQXTDAtGWpWAe4Plw

# The aim of this project is to use TensorFlow to transform our own data into TFRecord format.

# data: notMNIST
# http://yaroslavvb.blogspot.ca/2011/09/notmnist-dataset.html
# http://yaroslavvb.com/upload/notMNIST/
# 注意 ：文件下载地址如上，而非在代码里下载

# 导库
import tensorflow as tf
import numpy as np
import os
import matplotlib.pyplot as plt
import skimage.io as io   # anaconda 自带


# %% 通过数据文件地址得到images and labels
def get_file(file_dir):
    '''Get full image directory and corresponding labels
    Args:
        file_dir: file directory
    Returns:
        images: image directories, list, string
        labels: label, list, int
    '''

    images = []
    temp = []
    for root, sub_folders, files in os.walk(file_dir):
        # image directories
        for name in files:
            images.append(os.path.join(root, name))
        # get 10 sub-folder names
        for name in sub_folders:
            temp.append(os.path.join(root, name))

    # assign 10 labels based on the folder names
    labels = []
    for one_folder in temp:
        n_img = len(os.listdir(one_folder))
        letter = one_folder.split('/')[-1]
        # 通过子文件夹名字获取标签，标签编号为1-10

        if letter == 'A':
            labels = np.append(labels, n_img * [1])
        elif letter == 'B':
            labels = np.append(labels, n_img * [2])
        elif letter == 'C':
            labels = np.append(labels, n_img * [3])
        elif letter == 'D':
            labels = np.append(labels, n_img * [4])
        elif letter == 'E':
            labels = np.append(labels, n_img * [5])
        elif letter == 'F':
            labels = np.append(labels, n_img * [6])
        elif letter == 'G':
            labels = np.append(labels, n_img * [7])
        elif letter == 'H':
            labels = np.append(labels, n_img * [8])
        elif letter == 'I':
            labels = np.append(labels, n_img * [9])
        else:
            labels = np.append(labels, n_img * [10])

    # shuffle
    temp = np.array([images, labels])
    temp = temp.transpose()
    np.random.shuffle(temp)

    image_list = list(temp[:, 0])
    label_list = list(temp[:, 1])
    label_list = [int(float(i)) for i in label_list]

    return image_list, label_list

#%%  将标签和图像转为特定的格式：
# 因为后面convert_to_tfrecord函数用到的编码函数tf.train.Example的数据结构中包含了一个从属性到取值的字典。而属性的取值可以为
# 字符串列表(BytesList)、实数列表(FloatList)或者整数列表(Int64List)通过以下函数编码为Example proto形式的返回值，所以提前在这里转换
# 以下两个函数来自tensorflow官网
def int64_feature(value):
  """Wrapper for inserting int64 features into Example proto."""
  if not isinstance(value, list):
    value = [value]
  return tf.train.Feature(int64_list=tf.train.Int64List(value=value))

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


# %% 转换成TFRecord
def convert_to_tfrecord(images, labels, save_dir, name):
    '''convert all images and labels to one tfrecord file.
    Args:
        images: list of image directories, string type
        labels: list of labels, int type
        save_dir: the directory to save tfrecord file, e.g.: '/home/folder1/'
        name: the name of tfrecord file, string type, e.g.: 'train'
    Return:
        no return
    Note:
        converting needs some time, be patient...
        转换时间有点长
    '''

    # 指定数据转换格式后的保存路径和名称
    filename = os.path.join(save_dir, name + '.tfrecords')
    # 取得图片的样本总数
    n_samples = len(labels)

    # 检查图像和label是不是对应的
    if np.shape(images)[0] != n_samples:
        raise ValueError('Images size %d does not match label size %d.' % (images.shape[0], n_samples))

    # wait some time here, transforming need some time based on the size of your data.
    # 创建一个实例对象 writer，用于后面序列化数据的写入
    writer = tf.python_io.TFRecordWriter(filename)
    print('\nTransform start......')
    # 将所有数据(包括标签等)按照 tf.train.Example Protocol Buffer 的格式存储
    for i in np.arange(0, n_samples):
        try:
            image = io.imread(images[i])  # type(image) must be array! 一张一张的读取
            image_raw = image.tostring()   # 将图片矩阵转化为字符串
            label = int(labels[i])

            # 创建tf.train.Example 协议内存块，把标签、图片数据作为特定字段存入（数据类型转换）
            example = tf.train.Example(features=tf.train.Features(feature={
                'label': int64_feature(label),
                'image_raw': bytes_feature(image_raw)}))

            # 调用实例对象 writer 的 write 方法将序列化后的 example 协议内存块写入 TFRecord 文件
            writer.write(example.SerializeToString())
        except IOError as e:
            print('Could not read:', images[i])
            print('error: %s' % e)
            print('Skip it!\n')
            # 原始数据有几张图片有错误不能读取，用这种方式略过
    # 调用实例对象 writer 的 close 方法结束写入过程
    writer.close()
    print('Transform done!')


# %% 读取 、解码生成batch
# 以上将数据转换成了TFRecord形式，将其解码出来
def read_and_decode(tfrecords_file, batch_size):
    '''read and decode tfrecord file, generate (image, label) batches
    Args:
        tfrecords_file: the directory of tfrecord file
        batch_size: number of images in each batch
    Returns:
        image: 4D tensor - [batch_size, width, height, channel]
        label: 1D tensor - [batch_size]
    '''
    # make an input queue from the tfrecord file 创建文件名队列，方便数据的读取并节约时间
    # 关于队列的原理见知乎地址： https://zhuanlan.zhihu.com/p/27238630
    filename_queue = tf.train.string_input_producer([tfrecords_file])

    # 创建一个实例对象 reader, 用于读取 TFRecord 中的样例
    reader = tf.TFRecordReader()
    _, serialized_example = reader.read(filename_queue)

    # 按照字段格式，解析数据，从TFRecord 到 tf.int64、tf.string
    img_features = tf.parse_single_example(
        serialized_example,
        features={
            'label': tf.FixedLenFeature([], tf.int64),
            'image_raw': tf.FixedLenFeature([], tf.string),
        })
    # 解码
    image = tf.decode_raw(img_features['image_raw'], tf.uint8)

    ##########################################################
    # you can put data augmentation here, I didn't use it
    ##########################################################
    # all the images of notMNIST are 28*28, you need to change the image size if you use other dataset.

    image = tf.reshape(image, [28, 28])
    label = tf.cast(img_features['label'], tf.int32)
    image_batch, label_batch = tf.train.batch([image, label],
                                              batch_size=batch_size,
                                              num_threads=64,
                                              capacity=2000)
    # num_threads：可以指定多个线程同时执行入队操作(数据读取和预处理)，通过队列实现多线程处理机制
    # capacity: 队列中最多可以存储的样例个数
    return image_batch, tf.reshape(label_batch, [batch_size])

## 测试
# %% Convert data to TFRecord

test_dir = 'D:/python/neural network/notMINST/notMNIST_small/'
save_dir = 'D:/python/neural network/notMINST/'
BATCH_SIZE = 25

# Convert test data: you just need to run it ONCE !
name_test = 'test'
images, labels = get_file(test_dir)
convert_to_tfrecord(images, labels, save_dir, name_test)


# %% TO test train.tfrecord file

def plot_images(images, labels):
    '''plot one batch size
    '''
    for i in np.arange(0, BATCH_SIZE):
        plt.subplot(5, 5, i + 1)
        plt.axis('off')
        plt.title(chr(ord('A') + labels[i] - 1), fontsize=14)
        plt.subplots_adjust(top=1.5)
        plt.imshow(images[i])
    plt.show()


tfrecords_file = 'D:/python/neural network/notMINST/test.tfrecords'
# 这个地址为 save_file的地址加上数据的名字，详情见convert_to_tfrecord函数
image_batch, label_batch = read_and_decode(tfrecords_file, batch_size=BATCH_SIZE)

with tf.Session()  as sess:
    i = 0
    coord = tf.train.Coordinator()
    threads = tf.train.start_queue_runners(coord=coord)

    try:
        while not coord.should_stop() and i < 1:
            # just plot one batch size
            image, label = sess.run([image_batch, label_batch])
            plot_images(image, label)
            i += 1
    except tf.errors.OutOfRangeError:
        print('done!')
    finally:
        coord.request_stop()
    coord.join(threads)