tf. nn. top_ k()

Reference resources  tf.nn.top_k() - cloud + Community - Tencent cloud

One 、 The function prototype

tf.nn.top_k(
    input,
    k=1,
    sorted=True,
    name=None
)

To find the largest of the last dimension of the input tensor k A value and its subscript ！

If you enter a vector , That is to say rank=1, Find the biggest k The number is in this vector , Then the maximum k A number and the largest one k Subscript of a number . If the input tensor is a higher rank Matrix , Then we just need to find the largest in each line k A digital , And their subscripts . If two elements are the same , So the lower subscript appears first .

Parameters ：

• input: Input tensor, It can't be array These ！ Either type 1-D, For higher dimensions, the length of the last dimension must be greater than or equal to K
• k:0-D Of int32 Digital tensor of .
• sorted: If sorted=True, So the selected one k Numbers need to be sorted in descending order
• name: optional , This is the name of the operation

return ：

• values： That is, the largest of each line k A digital
• indices： The subscript here is the subscript in the last dimension of the input tensor

Two 、 Example

import tensorflow as tf
import numpy as np
     
# Select the largest first two digits of each line 
# The return is the largest k A digital , At the same time, the maximum k Numbers in the subscript of the last dimension 
a=tf.constant(np.random.rand(3,4))
b=tf.nn.top_k(a,k=2)
with tf.Session() as sess:
    print(sess.run(a))
    print(sess.run(b))


Output:

    [[0.73731748 0.13455566 0.20236765 0.92909052]
     [0.7923021  0.46949081 0.31521194 0.2999236 ]
     [0.19102823 0.01301476 0.70615716 0.68501807]]
    TopKV2(values=array([[0.92909052, 0.73731748],
           [0.7923021 , 0.46949081],
           [0.70615716, 0.68501807]]), indices=array([[3, 0],
           [0, 1],
           [2, 3]], dtype=int32))