message Example { 
  Features features = 1; 
}; 
message Features { 
  map<string, Feature> feature = 1; 
}; 
message Feature { 
  oneof kind { 
  BytesList bytes_list = 1; 
  FloatList float_list = 2; 
  Int64List int64_list = 3; 
} 
}; 

from random import shuffle 
import numpy as np 
import glob 
import tensorflow as tf 
import cv2 
import sys 
import os 
 
# 因為我裝的是CPU版本的，運行起來會有'warning'，解決方法入下，眼不見為凈~ 
os.environ['TF_CPP_MIN_LOG_LEVEL'] = '2' 
 
shuffle_data = True 
image_path = '/path/to/image/*.jpg' 
 
# 取得該路徑下所有圖片的路徑，type（addrs）= list 
addrs = glob.glob(image_path) 
# 標(biāo)簽數(shù)據(jù)的獲得具體情況具體分析，type（labels）= list 
labels = ... 
 
# 這里是打亂數(shù)據(jù)的順序 
if shuffle_data: 
  c = list(zip(addrs, labels)) 
  shuffle(c) 
  addrs, labels = zip(*c) 
 
# 按需分割數(shù)據(jù)集 
train_addrs = addrs[0:int(0.7*len(addrs))] 
train_labels = labels[0:int(0.7*len(labels))] 
 
val_addrs = addrs[int(0.7*len(addrs)):int(0.9*len(addrs))] 
val_labels = labels[int(0.7*len(labels)):int(0.9*len(labels))] 
 
test_addrs = addrs[int(0.9*len(addrs)):] 
test_labels = labels[int(0.9*len(labels)):] 
 
# 上面不是獲得了image的地址么，下面這個函數(shù)就是根據(jù)地址獲取圖片 
def load_image(addr): # A function to Load image 
  img = cv2.imread(addr) 
  img = cv2.resize(img, (224, 224), interpolation=cv2.INTER_CUBIC) 
  img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB) 
  # 這里/255是為了將像素值歸一化到[0，1] 
  img = img / 255. 
  img = img.astype(np.float32) 
  return img 
 
# 將數(shù)據(jù)轉(zhuǎn)化成對應(yīng)的屬性 
def _int64_feature(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])) 
 
 
def _float_feature(value): 
  return tf.train.Feature(float_list=tf.train.FloatList(value=[value])) 
 
# 下面這段就開始把數(shù)據(jù)寫入TFRecods文件 
 
train_filename = '/path/to/train.tfrecords' # 輸出文件地址 
 
# 創(chuàng)建一個writer來寫 TFRecords 文件 
writer = tf.python_io.TFRecordWriter(train_filename) 
 
for i in range(len(train_addrs)): 
  # 這是寫入操作可視化處理 
  if not i % 1000: 
    print('Train data: {}/{}'.format(i, len(train_addrs))) 
    sys.stdout.flush() 
  # 加載圖片 
  img = load_image(train_addrs[i]) 
 
  label = train_labels[i] 
 
  # 創(chuàng)建一個屬性（feature） 
  feature = {'train/label': _int64_feature(label), 
        'train/image': _bytes_feature(tf.compat.as_bytes(img.tostring()))} 
 
  # 創(chuàng)建一個 example protocol buffer 
  example = tf.train.Example(features=tf.train.Features(feature=feature)) 
 
  # 將上面的example protocol buffer寫入文件 
  writer.write(example.SerializeToString()) 
 
writer.close() 
sys.stdout.flush() 

import tensorflow as tf 
import numpy as np 
import matplotlib.pyplot as plt 
import os  
os.environ['TF_CPP_MIN_LOG_LEVEL'] = '2' 
data_path = 'train.tfrecords' # tfrecords 文件的地址 
 
with tf.Session() as sess: 
  # 先定義feature，這里要和之前創(chuàng)建的時候保持一致 
  feature = { 
    'train/image': tf.FixedLenFeature([], tf.string), 
    'train/label': tf.FixedLenFeature([], tf.int64) 
  } 
  # 創(chuàng)建一個隊列來維護輸入文件列表 
  filename_queue = tf.train.string_input_producer([data_path], num_epochs=1) 
 
  # 定義一個 reader ，讀取下一個 record 
  reader = tf.TFRecordReader() 
  _, serialized_example = reader.read(filename_queue) 
 
  # 解析讀入的一個record 
  features = tf.parse_single_example(serialized_example, features=feature) 
 
  # 將字符串解析成圖像對應(yīng)的像素組 
  image = tf.decode_raw(features['train/image'], tf.float32) 
 
  # 將標(biāo)簽轉(zhuǎn)化成int32 
  label = tf.cast(features['train/label'], tf.int32) 
 
  # 這里將圖片還原成原來的維度 
  image = tf.reshape(image, [224, 224, 3]) 
 
  # 你還可以進行其他一些預(yù)處理.... 
 
  # 這里是創(chuàng)建順序隨機 batches(函數(shù)不懂的自行百度) 
  images, labels = tf.train.shuffle_batch([image, label], batch_size=10, capacity=30, min_after_dequeue=10) 
 
  # 初始化 
  init_op = tf.group(tf.global_variables_initializer(), tf.local_variables_initializer()) 
  sess.run(init_op) 
 
  # 啟動多線程處理輸入數(shù)據(jù) 
  coord = tf.train.Coordinator() 
  threads = tf.train.start_queue_runners(coord=coord) 
 
  .... 
 
  #關(guān)閉線程 
  coord.request_stop() 
  coord.join(threads) 
  sess.close()