资源经验分享TensorFlow2.0(1):基本数据结构——张量

TensorFlow2.0(1):基本数据结构——张量

2019-12-10 | |  59 |   0

原标题:TensorFlow2.0(1):基本数据结构——张量

原文来自:博客园      原文链接:https://www.cnblogs.com/chenhuabin/p/11561633.html


阅读目录


 

 


 


注:本系列所有博客将持续更新并发布在github上,您可以通过github下载本系列所有文章笔记文件。


1引言

TensorFlow2.0版本已经发布,虽然不是正式版,但预览版都发布了,正式版还会远吗?相比于1.X,2.0版的TensorFlow修改的不是一点半点,这些修改极大的弥补了1.X版本的反人类设计,提升了框架的整体易用性,绝对好评!

赶紧来学习一波吧,做最先吃螃蟹的那一批人!先从TensorFlow的基本数据结构——张量(tensor)开始。


2 创建


2.1 constant()方法

In [1]:

import tensorflow as tf

In [2]:

tf.constant(1)  # 创建一个整型张量

Out[2]:

<tf.Tensor: id=0, shape=(), dtype=int32, numpy=1>

In [3]:

tf.constant(1.)  # 创建一个浮点型张量

Out[3]:

<tf.Tensor: id=2, shape=(), dtype=float32, numpy=1.0>

In [4]:

tf.constant(2., dtype=tf.double)  # 创建的同时指定数据类型

Out[4]:

<tf.Tensor: id=4, shape=(), dtype=float64, numpy=2.0>

In [5]:

tf.constant([[1.,2.,3.],[4.,5.,6.]])  # 通过传入一个list参数创建

Out[5]:

<tf.Tensor: id=6, shape=(2, 3), dtype=float32, numpy=
array([[1., 2., 3.],
       [4., 5., 6.]], dtype=float32)>

如果输入的数据与指定的数据类型不相符,会产生以下异常:
TypeError: Cannot convert provided value to EagerTensor. Provided value: 2.1 Requested dtype: int32


2.2 convert_to_tensor()方法

In [9]:

import numpy as np

In [10]:

tf.convert_to_tensor(np.ones([2, 3]))

Out[10]:

<tf.Tensor: id=9, shape=(2, 3), dtype=float64, numpy=
array([[1., 1., 1.],
       [1., 1., 1.]])>

In [11]:

tf.convert_to_tensor(np.ones([2, 3]))

Out[11]:

<tf.Tensor: id=11, shape=(2, 3), dtype=float64, numpy=
array([[1., 1., 1.],
       [1., 1., 1.]])>

In [12]:

tf.convert_to_tensor([[2.,3.],[3., 4.]])

Out[12]:

<tf.Tensor: id=13, shape=(2, 2), dtype=float32, numpy=
array([[2., 3.],
       [3., 4.]], dtype=float32)>


2.3 创建元素为指定值的tensor

如果你熟悉numpy创建数组的方法,你一定见过zeros()、ones()等方法,TensorFlow中也有这些方法。

(1)zeros()与ones()

In [24]:

a = tf.zeros([2, 3, 3])  # 创建一个元素全为0,形状为[2, 3, 3]的tensor

In [25]:

a

Out[25]:

<tf.Tensor: id=46, shape=(2, 3, 3), dtype=float32, numpy=
array([[[0., 0., 0.],
        [0., 0., 0.],
        [0., 0., 0.]],

       [[0., 0., 0.],
        [0., 0., 0.],
        [0., 0., 0.]]], dtype=float32)>

In [26]:

b = tf.ones([2, 3])  #  创建一个元素全为1,形状为[2, 3]的tensor

In [27]:

b

Out[27]:

<tf.Tensor: id=50, shape=(2, 3), dtype=float32, numpy=
array([[1., 1., 1.],
       [1., 1., 1.]], dtype=float32)>

(2)zeros_like()与ones_like

In [28]:

tf.zeros_like(b)  # 仿照b的shape创建一个全为0的tensor

Out[28]:

<tf.Tensor: id=52, shape=(2, 3), dtype=float32, numpy=
array([[0., 0., 0.],
       [0., 0., 0.]], dtype=float32)>

In [29]:

tf.ones_like(a)  # 仿照b的shape创建一个全为1的tensor

Out[29]:

<tf.Tensor: id=56, shape=(2, 3, 3), dtype=float32, numpy=
array([[[1., 1., 1.],
        [1., 1., 1.],
        [1., 1., 1.]],

       [[1., 1., 1.],
        [1., 1., 1.],
        [1., 1., 1.]]], dtype=float32)>

(3)fill()

In [21]:

tf.fill([2,3],5)  # 创建元素全为5,形状为[2,3]的tensor

Out[21]:

<tf.Tensor: id=38, shape=(2, 3), dtype=int32, numpy=
array([[5, 5, 5],
       [5, 5, 5]])>


2.4 随机初始化

在实际应用中,经常需要随机初始化元素服从某种分布的tensor,TensorFlow中也提供了这种功能。

(1)从指定正态分布中随机取值:tf.random.normal()。例如,随机初始化一个元素服从均值为1,方差为1的正态分布且形状为[2, 3]的tensor:

In [30]:

tf.random.normal([2, 3], mean=1, stddev=1)

Out[30]:

<tf.Tensor: id=63, shape=(2, 3), dtype=float32, numpy=
array([[ 1.7034731 ,  0.4979009 ,  1.4266468 ],
       [-0.33414853,  0.2618034 ,  0.3966313 ]], dtype=float32)>

(2)从指定的截断正态分布中随机取值:truncated_normal()。意思是从指定的正太分布中取值,但是取值范围在两个标准差范围内,也就是:[ mean - 2 stddev, mean + 2 stddev ]

In [31]:

tf.random.truncated_normal([2, 3], mean=1, stddev=1)

Out[31]:

<tf.Tensor: id=70, shape=(2, 3), dtype=float32, numpy=
array([[0.71736836, 1.7930655 , 0.47575486],
       [0.83504593, 0.7969478 , 0.6002228 ]], dtype=float32)>

(3)从指定均匀分布中随机取值:tf.random.uniform()。

In [32]:

tf.random.uniform([2, 3], minval=1, maxval=2) # 在1~2之间均匀分布

Out[32]:

<tf.Tensor: id=78, shape=(2, 3), dtype=float32, numpy=
array([[1.7117869, 1.2625391, 1.6652637],
       [1.3810604, 1.0297629, 1.1268978]], dtype=float32)>


3 索引

In [33]:

a = tf.convert_to_tensor(np.arange(80).reshape(2,2,4,5))

In [34]:

a

Out[34]:

<tf.Tensor: id=80, shape=(2, 2, 4, 5), dtype=int32, numpy=
array([[[[ 0,  1,  2,  3,  4],
         [ 5,  6,  7,  8,  9],
         [10, 11, 12, 13, 14],
         [15, 16, 17, 18, 19]],

        [[20, 21, 22, 23, 24],
         [25, 26, 27, 28, 29],
         [30, 31, 32, 33, 34],
         [35, 36, 37, 38, 39]]],


       [[[40, 41, 42, 43, 44],
         [45, 46, 47, 48, 49],
         [50, 51, 52, 53, 54],
         [55, 56, 57, 58, 59]],

        [[60, 61, 62, 63, 64],
         [65, 66, 67, 68, 69],
         [70, 71, 72, 73, 74],
         [75, 76, 77, 78, 79]]]])>


3.1 基础索引

TensorFlow支持Python原生的基础索引方式,即多个方括号逐步索引取值:[idx][idx][idx],每个方括号对应一个维度。

In [35]:

a[0]  # 取第一个维度

Out[35]:

<tf.Tensor: id=85, shape=(2, 4, 5), dtype=int32, numpy=
array([[[ 0,  1,  2,  3,  4],
        [ 5,  6,  7,  8,  9],
        [10, 11, 12, 13, 14],
        [15, 16, 17, 18, 19]],

       [[20, 21, 22, 23, 24],
        [25, 26, 27, 28, 29],
        [30, 31, 32, 33, 34],
        [35, 36, 37, 38, 39]]])>

In [36]:

a[0][1]  # 同时筛选两个维度

Out[36]:

<tf.Tensor: id=94, shape=(4, 5), dtype=int32, numpy=
array([[20, 21, 22, 23, 24],
       [25, 26, 27, 28, 29],
       [30, 31, 32, 33, 34],
       [35, 36, 37, 38, 39]])>

In [37]:

a[0][1][3][3]  # 同时对4个维度进行筛选

Out[37]:

<tf.Tensor: id=111, shape=(), dtype=int32, numpy=38>

这种索引数据的方法简单,易于理解,但是可读性差,只能按维度依次索引数据,也不能索引列。


3.2 numpy索引

TensorFlow也继承了numpy中的部分索引方式,如果对numpy索引方式不熟悉,可以查看我的前几篇博客
(1)[idx1, idx2, idx3]
这种索引方式是在一个方括号内写下所有的索引,每个索引序号之间用逗号隔开。

In [38]:

a[1]  # 筛选第一维度,这跟基础索引一样

Out[38]:

<tf.Tensor: id=116, shape=(2, 4, 5), dtype=int32, numpy=
array([[[40, 41, 42, 43, 44],
        [45, 46, 47, 48, 49],
        [50, 51, 52, 53, 54],
        [55, 56, 57, 58, 59]],

       [[60, 61, 62, 63, 64],
        [65, 66, 67, 68, 69],
        [70, 71, 72, 73, 74],
        [75, 76, 77, 78, 79]]])>

In [39]:

a[1,1, 3]  # 同时帅选3个维度

Out[39]:

<tf.Tensor: id=121, shape=(5,), dtype=int32, numpy=array([75, 76, 77, 78, 79])>

(2)冒号切片与步长:[start:end:step]

这种索引方式在Python原生的list类型中也是常见的,而且使用方法也是一样的。

In [40]:

a[1,:,0:2] # 对第1维度选第二块数据,对第二维度选所有数据,对第三维度选前两行

Out[40]:

<tf.Tensor: id=126, shape=(2, 2, 5), dtype=int32, numpy=
array([[[40, 41, 42, 43, 44],
        [45, 46, 47, 48, 49]],

       [[60, 61, 62, 63, 64],
        [65, 66, 67, 68, 69]]])>

In [41]:

a[1,:,0:2,0:4] # 继续上面的例子,对第4维度筛选去前4列

Out[41]:

<tf.Tensor: id=131, shape=(2, 2, 4), dtype=int32, numpy=
array([[[40, 41, 42, 43],
        [45, 46, 47, 48]],

       [[60, 61, 62, 63],
        [65, 66, 67, 68]]])>

In [42]:

a[1,:,0:2,0:4:2] # 对第4维度加上步长,每隔一个数据取一次

Out[42]:

<tf.Tensor: id=136, shape=(2, 2, 2), dtype=int32, numpy=
array([[[40, 42],
        [45, 47]],

       [[60, 62],
        [65, 67]]])>

也可以使用负值步长表示逆序索引,但要注意,负数步长时,原本的[start : end : step]也要跟着编程[end : start : step]:

In [43]:

a[1,:,0:2,4:0:-1]

Out[43]:

<tf.Tensor: id=141, shape=(2, 2, 4), dtype=int32, numpy=
array([[[44, 43, 42, 41],
        [49, 48, 47, 46]],

       [[64, 63, 62, 61],
        [69, 68, 67, 66]]])>

In [44]:

a[1,:,0:2,4:0:-2]

Out[44]:

<tf.Tensor: id=146, shape=(2, 2, 2), dtype=int32, numpy=
array([[[44, 42],
        [49, 47]],

       [[64, 62],
        [69, 67]]])>

在numpy和TensorFlow中还有“..."(三个英文句号)的使用,“..."用于表示连续多个维度全选:

In [45]:

a[1,...,0:4] # 等同于a[1, : , : ,0:4]

Out[45]:

<tf.Tensor: id=151, shape=(2, 4, 4), dtype=int32, numpy=
array([[[40, 41, 42, 43],
        [45, 46, 47, 48],
        [50, 51, 52, 53],
        [55, 56, 57, 58]],

       [[60, 61, 62, 63],
        [65, 66, 67, 68],
        [70, 71, 72, 73],
        [75, 76, 77, 78]]])>

In [46]:

a[0,0,...] # 等同于a[0,0,:,:]

Out[46]:

<tf.Tensor: id=156, shape=(4, 5), dtype=int32, numpy=
array([[ 0,  1,  2,  3,  4],
       [ 5,  6,  7,  8,  9],
       [10, 11, 12, 13, 14],
       [15, 16, 17, 18, 19]])>


3.3 gather与gather_nd

gather与gather_nd是指TensorFlow通过gather()方法和gather_nd()方法提供的两种索引方式。在numpy中,可以通过嵌套list的方式来指定无规则的索引:

In [47]:

b = np.arange(20).reshape(4,5)

In [48]:

b[1, [0,3,4]] # 选取第2行的第1列、第4列、第5列

Out[48]:

array([5, 8, 9])

但是在TensorFlow中,这种索引方式并没有从numpy中继承下来,所以如果在Tensor中使用这种方式,会抛出以下异常:
TypeError: Only integers, slices (:), ellipsis (...), tf.newaxis (None) and scalar tf.int32/tf.int64 tensors are valid indices, got [0, 3, 4]

还好的是,在TensorFlow中通过gather()方法和gather_nd()方法提供了这种索引方法。

(1)gather()方法

In [54]:

tf.gather(b, axis=0, indices=[0, 2, 3]) # 选取第1行,第3行,第4行

Out[54]:

<tf.Tensor: id=163, shape=(3, 5), dtype=int32, numpy=
array([[ 0,  1,  2,  3,  4],
       [10, 11, 12, 13, 14],
       [15, 16, 17, 18, 19]])>

In [55]:

tf.gather(b, axis=1, indices=[0, 2, 3]) # 选取第1列,第3列,第4列

Out[55]:

<tf.Tensor: id=168, shape=(4, 3), dtype=int32, numpy=
array([[ 0,  2,  3],
       [ 5,  7,  8],
       [10, 12, 13],
       [15, 17, 18]])>

仔细观察上面gather()方法例子,可以发现,第一个参数时数据源,还有两个参数中,axis指的是将要的维度,indices指的是需要选取的序号。

(2)gather_nd()

gather()方法一次只能对一个维度进行索引,gather_nd()方法可以同时对多个维度进行索引。

In [56]:

tf.gather_nd(b, [[0, 2],[3, 3]]) # 选取第1行第3列的那个数据,和第4行第4列的数据

Out[56]:

<tf.Tensor: id=172, shape=(2,), dtype=int32, numpy=array([ 2, 18])>


3.5 条件索引

可以结合一些简单的逻辑运算符进行索引取值:

In [1]:

import tensorflow as tf

In [6]:

a = tf.random.uniform([3,3],minval=-10,maxval=10,dtype=tf.int32)a

Out[6]:

<tf.Tensor: id=17, shape=(3, 3), dtype=int32, numpy=
array([[-7, -9,  6],
       [-6, -5,  9],
       [ 8,  9,  4]])>

In [7]:

mask = a < 0mask

Out[7]:

<tf.Tensor: id=20, shape=(3, 3), dtype=bool, numpy=
array([[ True,  True, False],
       [ True,  True, False],
       [False, False, False]])>

可以看到,返回的是一个shape与a相同的tensor,在a小于零的位置是True,大于零的位置为False。进一步地,我们可以用boolwan_mask()方法直接取出符合条件的元素:

In [13]:

tf.boolean_mask(a,mask)

Out[13]:

<tf.Tensor: id=82, shape=(4,), dtype=int32, numpy=array([-7, -9, -6, -5])>

可以结合where()方法取出符合条件元素的索引:

In [14]:

m_index = tf.where(mask)m_index

Out[14]:

<tf.Tensor: id=84, shape=(4, 2), dtype=int64, numpy=
array([[0, 0],
       [0, 1],
       [1, 0],
       [1, 1]], dtype=int64)>

再使用之前说过的gather_nd()方法取值:

In [15]:

tf.gather_nd(a,m_index)

Out[15]:

<tf.Tensor: id=86, shape=(4,), dtype=int32, numpy=array([-7, -9, -6, -5])>

where()方法还有第二种用法——从两个tensor中取出符合条件的值,这时候where()方法必须接受3个参数:

In [26]:

condition = tf.random.uniform([3,3],minval=0,maxval=2,dtype=tf.int32)condition = tf.cast(condition, tf.bool)condition

Out[26]:

<tf.Tensor: id=124, shape=(3, 3), dtype=bool, numpy=
array([[ True,  True, False],
       [ True, False, False],
       [ True, False, False]])>

In [34]:

a = tf.range(1,10)a = tf.reshape(a,[3,3])a

Out[34]:

<tf.Tensor: id=162, shape=(3, 3), dtype=int32, numpy=
array([[1, 2, 3],
       [4, 5, 6],
       [7, 8, 9]])>

In [35]:

b = tf.range(-9,0)b = tf.reshape(b,[3,3])b

Out[35]:

<tf.Tensor: id=169, shape=(3, 3), dtype=int32, numpy=
array([[-9, -8, -7],
       [-6, -5, -4],
       [-3, -2, -1]])>

In [36]:

tf.where(condition, a, b)

Out[36]:

<tf.Tensor: id=171, shape=(3, 3), dtype=int32, numpy=
array([[ 1,  2, -7],
       [ 4, -5, -4],
       [ 7, -2, -1]])>

上面where()方法返回的结果在True的位置取值是a中对应位置元素的值,在False位置是b中对应元素的值。

回到顶部

4 维度变换


4.1 reshape()

numpy中的ndarray数组有个一reshape()方法,用来改变数组的shape,TensorFlow中的reshape()方法,功能也是一样的,不过TensorFlow中的reshape()没有绑定到tensor中:

In [58]:

a = tf.ones([2,3,4])

In [59]:

a.shape

Out[59]:

TensorShape([2, 3, 4])

In [60]:

a

Out[60]:

<tf.Tensor: id=176, shape=(2, 3, 4), dtype=float32, numpy=
array([[[1., 1., 1., 1.],
        [1., 1., 1., 1.],
        [1., 1., 1., 1.]],

       [[1., 1., 1., 1.],
        [1., 1., 1., 1.],
        [1., 1., 1., 1.]]], dtype=float32)>

In [61]:

b = tf.reshape(a, [2, 2, 6])

In [62]:

b.shape

Out[62]:

TensorShape([2, 2, 6])

In [64]:

b

Out[64]:

<tf.Tensor: id=179, shape=(2, 2, 6), dtype=float32, numpy=
array([[[1., 1., 1., 1., 1., 1.],
        [1., 1., 1., 1., 1., 1.]],

       [[1., 1., 1., 1., 1., 1.],
        [1., 1., 1., 1., 1., 1.]]], dtype=float32)>

In [65]:

c = tf.reshape(a, [3, 2, 4])

In [66]:

c

Out[66]:

<tf.Tensor: id=183, shape=(3, 2, 4), dtype=float32, numpy=
array([[[1., 1., 1., 1.],
        [1., 1., 1., 1.]],

       [[1., 1., 1., 1.],
        [1., 1., 1., 1.]],

       [[1., 1., 1., 1.],
        [1., 1., 1., 1.]]], dtype=float32)>

可以看到,在上面的例子中,通过reshape()方法可以很方便的改变tensor的形状,得到一个新的tensor,需要注意的是在进行维度变换时,数据的重量是不变的,上面的例子无论是[2,3,4], [2, 2, 6]还是[3, 2, 4]都对应总量24,如果对应不上,就会产生异常。


4.2 转置:transpose()

transpose()方法提供了一种类似于装置的操作:

In [75]:

a = tf.constant([[1,2,3],[4,5,6]])

In [76]:

a.shape

Out[76]:

TensorShape([2, 3])

In [77]:

b = tf.transpose(a)

In [78]:

b.shape

Out[78]:

TensorShape([3, 2])

In [79]:

b

Out[79]:

<tf.Tensor: id=192, shape=(3, 2), dtype=int32, numpy=
array([[1, 4],
       [2, 5],
       [3, 6]])>

在默认情况下,transpose()方法会将所有维度按逆序方式完全转置,当然也可以通过perm参数执行需要转置的维度:

In [80]:

a=tf.constant([[[1,2,3],[4,5,6]],[[7,8,9],[10,11,12]]])

In [81]:

a

Out[81]:

<tf.Tensor: id=194, shape=(2, 2, 3), dtype=int32, numpy=
array([[[ 1,  2,  3],
        [ 4,  5,  6]],

       [[ 7,  8,  9],
        [10, 11, 12]]])>

In [82]:

b = tf.transpose(a) # 不指定perm参数时,相当于tf.transpose(a, perm=[2, 1, 0])

In [83]:

b

Out[83]:

<tf.Tensor: id=197, shape=(3, 2, 2), dtype=int32, numpy=
array([[[ 1,  7],
        [ 4, 10]],

       [[ 2,  8],
        [ 5, 11]],

       [[ 3,  9],
        [ 6, 12]]])>

In [84]:

c = tf.transpose(a, perm=[2, 1, 0])

In [85]:

c

Out[85]:

<tf.Tensor: id=200, shape=(3, 2, 2), dtype=int32, numpy=
array([[[ 1,  7],
        [ 4, 10]],

       [[ 2,  8],
        [ 5, 11]],

       [[ 3,  9],
        [ 6, 12]]])>

In [86]:

d = tf.transpose(a, perm=[0, 2, 1]) # 第一个维度不做变换,对第二、第三维度进行转置

In [87]:

d

Out[87]:

<tf.Tensor: id=203, shape=(2, 3, 2), dtype=int32, numpy=
array([[[ 1,  4],
        [ 2,  5],
        [ 3,  6]],

       [[ 7, 10],
        [ 8, 11],
        [ 9, 12]]])>


4.3 添加维度:expand_dims()

In [88]:

a=tf.constant([[1,2,3],[4,5,6]])

In [89]:

a

Out[89]:

<tf.Tensor: id=205, shape=(2, 3), dtype=int32, numpy=
array([[1, 2, 3],
       [4, 5, 6]])>

In [90]:

tf.expand_dims(a, axis=0)

Out[90]:

<tf.Tensor: id=208, shape=(1, 2, 3), dtype=int32, numpy=
array([[[1, 2, 3],
        [4, 5, 6]]])>

In [91]:

tf.expand_dims(a, axis=1)

Out[91]:

<tf.Tensor: id=211, shape=(2, 1, 3), dtype=int32, numpy=
array([[[1, 2, 3]],

       [[4, 5, 6]]])>

In [92]:

tf.expand_dims(a, axis=-1)

Out[92]:

<tf.Tensor: id=214, shape=(2, 3, 1), dtype=int32, numpy=
array([[[1],
        [2],
        [3]],

       [[4],
        [5],
        [6]]])>

In [93]:

tf.expand_dims(a, axis=2)

Out[93]:

<tf.Tensor: id=217, shape=(2, 3, 1), dtype=int32, numpy=
array([[[1],
        [2],
        [3]],

       [[4],
        [5],
        [6]]])>

expand_dims()方法添加维度时,通过axis参数指定添加维度的位置,正数表示从前往后数,负数表示从后往前数。


4.4 压缩维度:squeeze()

squeeze()方法与expand_dims()方法作用刚好相反,其作用是删除张量中dim为1的维度:

In [94]:

a = tf.ones([1,3,1,2])

In [95]:

a

Out[95]:

<tf.Tensor: id=221, shape=(1, 3, 1, 2), dtype=float32, numpy=
array([[[[1., 1.]],

        [[1., 1.]],

        [[1., 1.]]]], dtype=float32)>

In [96]:

tf.squeeze(a)

Out[96]:

<tf.Tensor: id=223, shape=(3, 2), dtype=float32, numpy=
array([[1., 1.],
       [1., 1.],
       [1., 1.]], dtype=float32)>



免责声明:本文来自互联网新闻客户端自媒体,不代表本网的观点和立场。

合作及投稿邮箱:E-mail:editor@tusaishared.com


上一篇:sklearn贝叶斯分类器的案例

下一篇:TensorFlow2.0(2):数学运算

用户评价
全部评价

热门资源

  • Python 爬虫(二)...

    所谓爬虫就是模拟客户端发送网络请求,获取网络响...

  • TensorFlow从1到2...

    原文第四篇中,我们介绍了官方的入门案例MNIST,功...

  • TensorFlow从1到2...

    “回归”这个词,既是Regression算法的名称,也代表...

  • TensorFlow2.0(10...

    前面的博客中我们说过,在加载数据和预处理数据时...

  • 机器学习中的熵、...

    熵 (entropy) 这一词最初来源于热力学。1948年,克...