TensorFlow2-索引与切片(6)

索引与切片

Indexing 索引

• Basic indexing
  • [idx][idx][idx]
• Same with Numpy
• [idx, idx,…]
• start:end
• start🔚step
• …

import tensorflow as tf
import numpy as np

Basic indexing

• 索引方式通用
• 单一

a=tf.ones([1,5,5,3])

a.shape

TensorShape([1, 5, 5, 3])

a[0][0]

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

a[0][0][0]

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

a[0][0][0][2]

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

Numpy-style indexing

a=tf.random.normal([4,28,28,3])

a[1].shape

TensorShape([28, 28, 3])

a[1,2].shape

TensorShape([28, 3])

a[1,2,3].shape

TensorShape([3])

a[1,2,3]

<tf.Tensor: shape=(3,), dtype=float32, numpy=array([-0.1556891 , -0.03499719,  1.491767  ], dtype=float32)>

a[1,2,3,2].shape

TensorShape([])

a[1,2,3,2]

<tf.Tensor: shape=(), dtype=float32, numpy=1.491767>

start:end

a=tf.range(10)
a

<tf.Tensor: shape=(10,), dtype=int32, numpy=array([0, 1, 2, 3, 4, 5, 6, 7, 8, 9], dtype=int32)>

a[0]

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

a[-1:]

<tf.Tensor: shape=(1,), dtype=int32, numpy=array([9], dtype=int32)>

a[-2:]

<tf.Tensor: shape=(2,), dtype=int32, numpy=array([8, 9], dtype=int32)>

a[:2]

<tf.Tensor: shape=(2,), dtype=int32, numpy=array([0, 1], dtype=int32)>

a[:-1]

<tf.Tensor: shape=(9,), dtype=int32, numpy=array([0, 1, 2, 3, 4, 5, 6, 7, 8], dtype=int32)>

Indexing by 综合使用

a=tf.random.normal([4,28,28,3])

a.shape

TensorShape([4, 28, 28, 3])

后置切片

a[0].shape

TensorShape([28, 28, 3])

a[0,:,:,:].shape

TensorShape([28, 28, 3])

a[0,1,:,:].shape

TensorShape([28, 3])

a[0,1,1,:].shape

TensorShape([3])

前置切片

a[:,:,:,:].shape

TensorShape([4, 28, 28, 3])

a[:,:,:,0].shape

TensorShape([4, 28, 28])

a[:,:,:,2].shape

TensorShape([4, 28, 28])

前后切片

a[:,0,:,:].shape

TensorShape([4, 28, 3])

Indexing by ::

• start🔚step
• ::step

a=tf.random.normal([4,28,28,3])

a.shape

TensorShape([4, 28, 28, 3])

# 第一维度[0-2]的样本
a[0:2,:,:,:].shape

TensorShape([2, 28, 28, 3])

# 2、3维度[0-28]间隔2个取值
a[:,0:28:2,0:28:2,:].shape

TensorShape([4, 14, 14, 3])

# 2、3维度[0-14]取值
a[:,:14,:14,:].shape

TensorShape([4, 14, 14, 3])

# 2、3维度[14：]取值
a[:,14:,14:,:].shape

TensorShape([4, 14, 14, 3])

# 2、3维度[0-28]间隔2个取值
a[:,::2,::2,:].shape

TensorShape([4, 14, 14, 3])

::-1 倒序

a=tf.range(4)
a

<tf.Tensor: shape=(4,), dtype=int32, numpy=array([0, 1, 2, 3], dtype=int32)>

a[::-1]

<tf.Tensor: shape=(4,), dtype=int32, numpy=array([3, 2, 1, 0], dtype=int32)>

a[::-2]

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

a[2::-2]

<tf.Tensor: shape=(2,), dtype=int32, numpy=array([2, 0], dtype=int32)>

… 省略号的方式

a=tf.random.normal([2,4,28,28,3])

a.shape

TensorShape([2, 4, 28, 28, 3])

a[0].shape

TensorShape([4, 28, 28, 3])

a[0,:,:,:,:].shape

TensorShape([4, 28, 28, 3])

a[0,...].shape

TensorShape([4, 28, 28, 3])

a[0,:,...].shape

TensorShape([4, 28, 28, 3])

a[:,:,:,:,0].shape

TensorShape([2, 4, 28, 28])

a[...,0].shape

TensorShape([2, 4, 28, 28])

a[0,...,2].shape

TensorShape([4, 28, 28])

a[1,0,...,2].shape

TensorShape([28, 28])

Selective Indexing 选择性索引

• tf.gather
• tf.gather_nd
• tf.boolean_ma

tf.gather

• data: [classes, students, subjects]
  • [4, 35, 8]

a=tf.random.normal([4,35,8])

a.shape

TensorShape([4, 35, 8])

# axis：选择的维度
# indices：索引号
tf.gather(a, axis=0,indices=[2,3]).shape

TensorShape([2, 35, 8])

a[2:4].shape

TensorShape([2, 35, 8])

tf.gather(a, axis=0,indices=[2,1,3,0]).shape

TensorShape([4, 35, 8])

tf.gather(a, axis=1,indices=[2,3,7,9,16]).shape

TensorShape([4, 5, 8])

tf.gather(a, axis=2,indices=[2,3,7]).shape

TensorShape([4, 35, 3])

tf.gather_nd

• data: [classes, students, subjects]
• What if sample several students and their several subjects？如果抽样几个学生和他们的几个科目呢?
• 𝑎𝑎 = 𝑡𝑓. 𝑔𝑎𝑡ℎ𝑒𝑟(𝑎, 𝑎𝑥𝑖𝑠,[𝑠𝑒𝑣𝑒𝑟𝑎𝑙 𝑠𝑡𝑢𝑑𝑒𝑛𝑡𝑠])
• 𝑎𝑎𝑎 = 𝑡𝑓. 𝑔𝑎𝑡ℎ𝑒𝑟(𝑎𝑎, 𝑎𝑥𝑖𝑠,[𝑠𝑒𝑣𝑒𝑟𝑎𝑙 𝑠𝑢𝑏𝑗𝑒𝑐𝑡𝑠])

a=tf.random.normal([4,35,8])

a.shape

TensorShape([4, 35, 8])

aa=tf.gather(a,axis=1,indices=[1,3,5,7,8,12,16])

aa.shape

TensorShape([4, 7, 8])

aaa=tf.gather(aa,axis=2,indices=[0,1,2])

aaa.shape

TensorShape([4, 7, 3])

tf.gather_nd

• data: [classes, students, subjects]
• What if sample several (classes and students)？如果抽样几个（班级和学生）呢
• 𝑓𝑜𝑟 𝑖𝑛𝑠𝑡𝑎𝑛𝑐𝑒:
  • [class1_student1, class2_studnet2, class3_student3, class4_student4]
  • → [4, 8]

a=tf.random.normal([4,35,8])

a.shape

TensorShape([4, 35, 8])

# [[0]] 第0个班级
# a[0]  -> [35,8]
tf.gather_nd(a, [0]).shape

TensorShape([35, 8])

# 第0个班级，第1学生
# a[0,1]  -> [8]
# a[0][1] -> [8]
tf.gather_nd(a, [0,1]).shape

TensorShape([8])

# 第0个班级，第1学生，第2门课
# a[0,1,2]  -> []
# a[0][1][2] -> []
tf.gather_nd(a, [0,1,2]).shape

TensorShape([])

tf.gather_nd(a, [0,1,2])

<tf.Tensor: shape=(), dtype=float32, numpy=0.012457304>

tf.gather_nd(a, [[0,1,2]]).shape

TensorShape([1])

# a[0][1][2] -> 标量
# 返长度为1的数据
tf.gather_nd(a, [[0,1,2]])

<tf.Tensor: shape=(1,), dtype=float32, numpy=array([0.0124573], dtype=float32)>

更复杂的使用方式

# [a[0][0],a[1][1]]
tf.gather_nd(a, [[0,0],[1,1]]).shape

TensorShape([2, 8])

# [a[0][0],a[1][1],a[2][2]]
tf.gather_nd(a, [[0,0],[1,1],[2,2]]).shape

TensorShape([3, 8])

# [a[0][0][0],a[1][1][1],a[2][2][2]]
tf.gather_nd(a, [[0,0,0],[1,1,1],[2,2,2]]).shape

TensorShape([3])

# [[a[0][0][0],a[1][1][1],a[2][2][2]]]
tf.gather_nd(a, [[[0,0,0],[1,1,1],[2,2,2]]]).shape

TensorShape([1, 3])

tf.gather_nd

• recommended ind
• [[0], [1],…]
• [[0,0], [1,1],…]
• [[0,0,0], [1,1,1],…]
• 推荐2层的写法:[[0]]

tf.boolean_mask

a=tf.random.normal([4,28,28,3])

a.shape

TensorShape([4, 28, 28, 3])

tf.boolean_mask(a, mask=[True, True,False,False]).shape

TensorShape([2, 28, 28, 3])

tf.boolean_mask(a, mask=[True, True,False],axis=3).shape

TensorShape([4, 28, 28, 2])

a=tf.ones([2,3,4])

# [True,False,False]  第1维；第2维；第3维

tf.boolean_mask(a, mask=[[True,False,False],[False, True,True]])

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