生成TFrecord

需求：将图片文件保存成Tfrecord的格式.

解决方法:基于tensorflow、cv2、numpy等库完成该功能.

注：改编自网上代码 

1)  准备要训练的手写识别的图片文件，并按照目录结构存放。见下图示意：

2)  生成训练图片和标签对应的文本文件，见下图示意：

3） 编写图片生成TFrecord代码，代码见下：

import numpy as np
import cv2
import tensorflow as tf
import matplotlib.pyplot as plt
from io import StringIO,BytesIO


# 将value转化成int64字节属性
def _int64_feature(value):
    return tf.train.Feature(int64_list=tf.train.Int64List(value=[value]))


# 将value转化成bytes属性
def _bytes_feature(value):
    return tf.train.Feature(bytes_list=tf.train.BytesList(value=[value]))


# 从训练文本里读取样本、标签。返回样本、标签列表以及行数
def load_file(examples_list_file):
    # type: (object) -> object
    lines = np.genfromtxt(examples_list_file, delimiter=" ", dtype=[('col1', 'S120'), ('col2', 'i8')])
    examples = []
    labels = []
    for example, label in lines:
        examples.append(example)
        labels.append(label)
    return np.asarray(examples), np.asarray(labels), len(lines)




def extract_image(filename, resize_height, resize_width):
    image = cv2.imread(filename)
    image = cv2.resize(image, (resize_height, resize_width))
    b, g, r = cv2.split(image)
    rgb_image = cv2.merge([r, g, b])
    rgb_image = rgb_image / 255.
    rgb_image = rgb_image.astype(np.float32)
    return rgb_image


def Image2TFRecord(trainDir,trainLabelFile,tfFile):
    resize_height = 28
    resize_width = 28
    train_file_root = trainDir
    train_file = trainLabelFile
    examples, labels, examples_num = load_file(train_file)
    writer = tf.python_io.TFRecordWriter(tfFile)
    for i, [example, label] in enumerate(zip(examples, labels)):
        print('No.%d' % (i))
        print(examples[i].decode(encoding="utf-8"))
        root = train_file_root + '/' + examples[i].decode(encoding="utf-8")
        print(root)
        image = extract_image(root, resize_height, resize_width)
        a = image.shape
        print(root)
        print('shape: %d, %d, %d, label: %d' % (image.shape[0], image.shape[1], image.shape[2], label))
        image_raw = image.tostring()  # 将Image转化成字符
        example = tf.train.Example(features=tf.train.Features(feature={
            'image_raw': _bytes_feature(image_raw),
            'height': _int64_feature(image.shape[0]),
            'width': _int64_feature(image.shape[1]),
            'depth': _int64_feature(image.shape[2]),
            'label': _int64_feature(label)
        }))
        writer.write(example.SerializeToString())
    writer.close()


if __name__ == '__main__':
    Image2TFRecord('E:/Python/mnist_img_data4','E:/Python/mnist_img_data4/train.txt','E:/Python/mnist_img_output/a44.tfrecords')
    TFrcords2Img('E:/Python/mnist_img_output/a44.tfrecords')


# 延展阅读 https://docs.scipy.org/doc/numpy-1.10.1/user/basics.io.genfromtxt.html
    指定按照定长3来分割
    data = "  1  2  3\n  4  5 67\n890123  4"
    print(np.genfromtxt(StringIO(data), delimiter=3))
    ''' 结果：
[[  1.   2.   3.]
 [  4.   5.  67.]
 [890. 123.   4.]]


    '''
     指定按照3列来分割,第一列长度是4,第二列长度是3,第三列长度是2
    data = "123456789\n   4  7 9\n   4567 9"
    print(np.genfromtxt(StringIO(data), delimiter=(4, 3, 2)))


    '''结果：
    [[1234.  567.   89.]
 [   4.    7.    9.]
 [   4.  567.    9.]]
    '''


    data = "1, abc , 2\n 3, xxx, 4"
     默认是不自动替换空格的
    print(np.genfromtxt(StringIO(data), delimiter=",", dtype="|S5"))
    print(np.genfromtxt(StringIO(data), delimiter=",", dtype="|S5", autostrip=True))
'''
[[b'1' b' abc ' b' 2']
 [b'3' b' xxx' b' 4']]
 
[[b'1' b'abc' b'2']
 [b'3' b'xxx' b'4']]


'''
data = "1 2 3\n4 5 6"
print(np.genfromtxt(StringIO(data), usecols=(0, -1)))
'''
[[1. 3.]
 [4. 6.]]
'''
print(np.genfromtxt(StringIO(data),names="a, b, c", usecols=("b", "c")))
'''
[(2., 3.) (5., 6.)]
'''
DataType的类型: