TensorFlow2-基础（二）：Tensor常见阶Rank【Rank = 0阶[]、1阶[d]、2阶[h, w]、3阶[b, len, vec]、4阶[b, h, w, c] 、5阶（元学习）】

一、Rank

Tensorflow将张量定义为多维数组。它可以表示任意维度的数组，维度用rank（阶）表示。它在深度学习中存在的意义是创造更高维度的矩阵、向量。

import pandas as pd
import tensorflow as tf

a1 = [
    [[1, 2, 3],
     [4, 5, 6],
     [7, 8, 9],
     [10, 11, 12]],

    [[11, 21, 31],
     [41, 51, 61],
     [71, 81, 91],
     [101, 111, 121]]
]

a2 = [
    [1, 2, 3],
    [4, 5, 6]
]

t1 = tf.convert_to_tensor(a1, dtype=tf.int32)
t2 = tf.convert_to_tensor(a2, dtype=tf.int32)

print("tf.rank(t1) = {0}; t1 = \n{1}".format(tf.rank(t1), t1))
print("-" * 100)
print("tf.rank(t2) = {0}; t2 = \n{1}".format(tf.rank(t2), t2))

打印结果：

tf.rank(t1) = 3; t1 = 
[[[  1   2   3]
  [  4   5   6]
  [  7   8   9]
  [ 10  11  12]]

 [[ 11  21  31]
  [ 41  51  61]
  [ 71  81  91]
  [101 111 121]]]
----------------------------------------------------------------------------------------------------
tf.rank(t2) = 2; t2 = 
[[1 2 3]
 [4 5 6]]

Process finished with exit code 0

rank：就是有几个维度，也相当与有几个括号，也是shape元组的长度，rank的idx计数都是从外向里，从大到小，也就是shape的排列方式。

• tf.rank(t1) = 3
• tf.rank(t2) = 2

图中数字表示rank的idx，箭头表示idx的方向。

二、典型的rank

Typical Dim Data

• []
• [d]
• [h, w]
• [b, len, vec]
• [b, h, w, c]
• [t, b, h, w, c]
• …

0、rank = 0维【标量/Scalar】

• []
• 0, 1., 2.2…
• accuracy
• loss = mse(out, y)

1、rank = 1维【向量/Vector】

Bias

• [out_dim]
  
  

2、rank = 2维【矩阵/Matrix】

• input x: [b, vec_dim]
• weight: [input_dim, output_dim]
  
  

3、rank = 3维【矩阵/Matrix】

x: [b, seq_len, word_dim]

• [b, 5, 5]

4、rank = 4维【矩阵/Matrix】

Image: [b, h, w, 3]

feature maps: [b, h, w, c]

5、rank= 5维【矩阵/Matrix】

Single task: [b, h, w, 3]

meta-learning：[task_b, b, h, w, 3]

---

---

---

参考资料：
tensorflow：理解 rank, shape, type
通俗易懂的理解Tensor(张量)、Rank、Shape