tensorflow中数据批次划分示例教程

1.简介

将数据划分成若干批次的数据，可以使用tf.train或者tf.data.Dataset中的方法。

1.1 tf.train

tf.train.slice_input_producer(tensor_list,shuffle=True,seed=None,capacity=32)

tf.train.batch(tensors,batch_size,num_threads=1,capacity=32,allow_smaller_final_batch=False)

参数说明：

shuffle：为True时进行数据清洗

allow_smaller_final_batch：为True时将小于batch_size的批次值输出

-------------------------------------------------------------------------------------------------------------------------

-------------------------------------------------------------------------------------------------------------------------

1.2 tf.data.Dataset

tf.data.Dataset是一个类，可以使用以下方法：

from_tensor_slices(tensors)

batch(batch_size,drop_remainder=False)

shuffle(buffer_size,seed=None,reshuffle_each_iteration=None)

repeat(count=None)

make_one_shot_iterator() / get_next()

注：make_one_shot_iterator() / get_next()用于Dataset数据的迭代器

参数说明：

tensors：可以是列表、字典、元组等类型

drop_remainder：为False时表示不保留小于batch_size的批次，否则删除

buffer_size：数据清洗时使用的buffer大小

count：对应为epoch个数，为None时表示数据序列无限延续

2.示例

2.1 使用tf.train.slice_input_producer和tf.train.batch

import tensorflow as tf
import numpy as np
import math

# 生成样例数据集
def generate_data():
    num = 15
    labels = np.asarray(range(num))
    images = np.random.random([num, 5, 5, 3])
    return images, labels

# 打印样例信息
images, labels = generate_data()
print('images.shape={0}, labels.shape={1}'.format(images.shape, labels.shape))

# 定义周期、批次、数据总量和遍历一次所有数据所需的迭代次数
n_epochs = 3
batch_size = 6
train_nums = 15
iterations = math.ceil(train_nums/batch_size)

# 使用tf.train.slice_input_producer将所有数据放入队列，使用tf.train.batch划分队列中的数据
input_queue = tf.train.slice_input_producer([images, labels], shuffle=False)
image_batch, label_batch = tf.train.batch(input_queue, batch_size=batch_size, num_threads=1, capacity=32)
print('image_batch.shape={0}, label_batch.shape={1}'.format(image_batch.shape, label_batch.shape))


with tf.Session() as sess:
    tf.global_variables_initializer().run()
    # 启动队列线程
    coord = tf.train.Coordinator()
    threads = tf.train.start_queue_runners(sess, coord)
    # 打印信息
    for epoch in range(n_epochs):       
        for iteration in range(iterations):
            cu_image_batch, cu_label_batch = sess.run([image_batch, label_batch])
            print('The {0} epoch, the {1} iteration, current batch is {2}'.format(epoch+1,iteration+1,cu_label_batch))
    # 接收线程
    coord.request_stop()
    coord.join(threads)    


# 打印结果如下
images.shape=(15, 5, 5, 3), labels.shape=(15,)
image_batch.shape=(6, 5, 5, 3), label_batch.shape=(6,)
The 1 epoch, the 1 iteration, current batch is [0 1 2 3 4 5]
The 1 epoch, the 2 iteration, current batch is [ 6  7  8  9 10 11]
The 1 epoch, the 3 iteration, current batch is [12 13 14  0  1  2]
The 2 epoch, the 1 iteration, current batch is [3 4 5 6 7 8]
The 2 epoch, the 2 iteration, current batch is [ 9 10 11 12 13 14]
The 2 epoch, the 3 iteration, current batch is [0 1 2 3 4 5]
The 3 epoch, the 1 iteration, current batch is [ 6  7  8  9 10 11]
The 3 epoch, the 2 iteration, current batch is [12 13 14  0  1  2]
The 3 epoch, the 3 iteration, current batch is [3 4 5 6 7 8]

如果tf.train.slice_input_producer(shuffle=True)，输出为乱序，结果如下：

images.shape=(15, 5, 5, 3), labels.shape=(15,)
image_batch.shape=(6, 5, 5, 3), label_batch.shape=(6,)
The 1 epoch, the 1 iteration, current batch is [ 2  5  8 11  3 10]
The 1 epoch, the 2 iteration, current batch is [ 9 12  7  1 14 13]
The 1 epoch, the 3 iteration, current batch is [0 6 4 2 3 6]
The 2 epoch, the 1 iteration, current batch is [11 10 12 14 13  5]
The 2 epoch, the 2 iteration, current batch is [8 1 0 9 4 7]
The 2 epoch, the 3 iteration, current batch is [10 13  1  4 12  3]
The 3 epoch, the 1 iteration, current batch is [ 2  8  5  9 14  7]
The 3 epoch, the 2 iteration, current batch is [ 0 11  6  1 14  9]
The 3 epoch, the 3 iteration, current batch is [11  6 12  7  0 13]

如果tf.train.batch(allow_smaller_final_batch=True)，则会返回不足批次数目的数据，结果如下：

images.shape=(15, 5, 5, 3), labels.shape=(15,)
The 1 epoch, the 1 iteration, current batch is [0 1 2 3 4 5]
The 1 epoch, the 2 iteration, current batch is [ 6  7  8  9 10 11]
The 1 epoch, the 3 iteration, current batch is [12 13 14]
The 2 epoch, the 1 iteration, current batch is [0 1 2 3 4 5]
The 2 epoch, the 2 iteration, current batch is [ 6  7  8  9 10 11]
The 2 epoch, the 3 iteration, current batch is [12 13 14]
The 3 epoch, the 1 iteration, current batch is [0 1 2 3 4 5]
The 3 epoch, the 2 iteration, current batch is [ 6  7  8  9 10 11]
The 3 epoch, the 3 iteration, current batch is [12 13 14]

2.2 使用tf.data.Dataset类

import tensorflow as tf
import numpy as np
import math

# 生成样例数据集
def generate_data():
    num = 15
    labels = np.asarray(range(num))
    images = np.random.random([num, 5, 5, 3])
    return images, labels
# 打印样例信息
images, labels = generate_data()
print('images.shape={0}, labels.shape={1}'.format(images.shape, labels.shape))

# 定义周期、批次、数据总数、遍历一次所有数据需的迭代次数
n_epochs = 3
batch_size = 6
train_nums = 15
iterations = math.ceil(train_nums/batch_size)

# 使用from_tensor_slices将数据放入队列，使用batch和repeat划分数据批次，且让数据序列无限延续
dataset = tf.data.Dataset.from_tensor_slices((images, labels))
dataset = dataset.batch(batch_size).repeat()

# 使用生成器make_one_shot_iterator和get_next取数据
iterator = dataset.make_one_shot_iterator()
next_iterator = iterator.get_next()

with tf.Session() as sess:
    for epoch in range(n_epochs):
        for iteration in range(iterations):
            cu_image_batch, cu_label_batch = sess.run(next_iterator)
            print('The {0} epoch, the {1} iteration, current batch is {2}'.format(epoch+1,iteration+1,cu_label_batch))


# 结果如下：
images.shape=(15, 5, 5, 3), labels.shape=(15,)
The 1 epoch, the 1 iteration, current batch is [0 1 2 3 4 5]
The 1 epoch, the 2 iteration, current batch is [ 6  7  8  9 10 11]
The 1 epoch, the 3 iteration, current batch is [12 13 14]
The 2 epoch, the 1 iteration, current batch is [0 1 2 3 4 5]
The 2 epoch, the 2 iteration, current batch is [ 6  7  8  9 10 11]
The 2 epoch, the 3 iteration, current batch is [12 13 14]
The 3 epoch, the 1 iteration, current batch is [0 1 2 3 4 5]
The 3 epoch, the 2 iteration, current batch is [ 6  7  8  9 10 11]
The 3 epoch, the 3 iteration, current batch is [12 13 14]

使用shuffle()，第23行修改为dataset = dataset.shuffle(100).batch(batch_size).repeat()，结果如下：

images.shape=(15, 5, 5, 3), labels.shape=(15,)
The 1 epoch, the 1 iteration, current batch is [ 7  4 10  8  3 11]
The 1 epoch, the 2 iteration, current batch is [ 0  2 12 13 14  5]
The 1 epoch, the 3 iteration, current batch is [6 9 1]
The 2 epoch, the 1 iteration, current batch is [ 6 14  7  9  3  8]
The 2 epoch, the 2 iteration, current batch is [13  5 12  1 11  2]
The 2 epoch, the 3 iteration, current batch is [ 0  4 10]
The 3 epoch, the 1 iteration, current batch is [10  8 13 12  3 14]
The 3 epoch, the 2 iteration, current batch is [ 6  9  2  5  1 11]
The 3 epoch, the 3 iteration, current batch is [0 4 7]

！！！