import numpy as np
def batch_generator(all_data , batch_size, shuffle=True):
 """
 :param all_data : all_data整个数据集，包含输入和输出标签
 :param batch_size: batch_size表示每个batch的大小
 :param shuffle: 是否打乱顺序
 :return:
 """
 # 输入all_datas的每一项必须是numpy数组，保证后面能按p所示取值
 all_data = [np.array(d) for d in all_data]
 # 获取样本大小
 data_size = all_data[0].shape[0]
 print("data_size: ", data_size)
 if shuffle:
  # 随机生成打乱的索引
  p = np.random.permutation(data_size)
  # 重新组织数据
  all_data = [d[p] for d in all_data]
 batch_count = 0
 while True:
  # 数据一轮循环(epoch)完成，打乱一次顺序
  if batch_count * batch_size + batch_size > data_size:
   batch_count = 0
   if shuffle:
    p = np.random.permutation(data_size)
    all_data = [d[p] for d in all_data]
  start = batch_count * batch_size
  end = start + batch_size
  batch_count += 1
  yield [d[start: end] for d in all_data]

# 输入x表示有23个样本，每个样本有两个特征
# 输出y表示有23个标签，每个标签取值为0或1
x = np.random.random(size=[23, 2])
y = np.random.randint(2, size=[23,1])
count = x.shape[0]
batch_size = 5
epochs = 20
batch_num = count // batch_size
batch_gen = batch_generator([x, y], batch_size)
for i in range(epochs):
 print("##### epoch %s ##### " % i)
 for j in range(batch_num):
  batch_x, batch_y = next(batch_gen)
  print("-----epoch=%s, batch=%s-----" % (i, j))
  print(batch_x, batch_y)

import tensorflow as tf
import numpy as np
def _parse_function(x):
 num_list = np.arange(10)
 return num_list
def _from_tensor_slice(x):
 return tf.data.Dataset.from_tensor_slices(x)
softmax_data = tf.data.Dataset.range(1000) # 构造一个队列
softmax_data = softmax_data.map(lambda x:tf.py_func(_parse_function, [x], [tf.int32]))# 将数据进行传入
softmax_data = softmax_data.flat_map(_from_tensor_slice) #将数据进行平铺, 将其变为一维的数据，from_tensor_slice将数据可以输出
softmax_data = softmax_data.batch(1) #构造一个batch的数量
softmax_iter = softmax_data.make_initializable_iterator() # 构造数据迭代器
softmax_element = softmax_iter.get_next() # 获得一个batch的数据
sess = tf.Session()
sess.run(softmax_iter.initializer) # 数据迭代器的初始化操作
print(sess.run(softmax_element)) # 实际获得一个数据
print(sess.run(softmax_data))