1. Code

1.1 Source code (copy and use directly)

:star: Copy the following code directly to the model for use (it has been tested and can be discussed in the comment area)

import keras.backend as K

def my_tp_tn_fp_fn(y_true, y_pred):
    true_posi_sum = K.cast(K.sum(y_true), "int32")
    true_nag_sum = K.cast(K.sum(y_true-1), "int32")*(-1)
    pred_posi_sum = K.sum(K.cast(K.greater(y_pred, 0.5), "int32"))

    tp = K.sum(K.cast(K.greater(K.clip(y_true * y_pred, 0.0, 1.0), 0.50), "int32"))
    fn = true_posi_sum - tp
    fp = pred_posi_sum - tp
    tn = true_nag_sum - fp

    tp = K.cast(tp, "float32")
    tn = K.cast(tn, "float32")
    fp = K.cast(fp, "float32")
    fn = K.cast(fn, "float32")

    return tp, tn, fp, fn

def keras_hanmingloss(y_true, y_pred):
    tp, tn, fp, fn = my_tp_tn_fp_fn(y_true, y_pred)
    num_wrong = fp + fn
    total = tp + tn + fp + fn
    hanming_loss = (num_wrong + K.epsilon())/ total

    return hanming_loss
复制代码

1.2 Construction ideas

flow chart

flowchart TD
step1[建立函数求出TP_TN_FP_FN]
step2[根据汉明损失原理构造损失函数]
step3[ 结合Sci-kitlearn库的hanminglossAPI进行验证]

step1 --> step2 
step2 --> step3

First find all TP, TN, FP, FN
Then construct the loss function according to all TP, TN, FP, FN combined with the Hamming loss formula
Validate with the hanmingloss API of the Sci-kitlearn library

1.3 Verification with sklearn

The custom code effect, as shown in the figure below, is:0.06353355

The Hamming loss effect of sklearn, as shown in the figure below, is: 0.06353354978354979(Because the former is float32, the latter float64 precision will be higher)

在这里插入图片描述

1.4 Experimental effect

The effect of the training phase

Evaluate Phase Effects

2. Hamming loss

2.1 Introduction

Hamming distance: find the number of different elements in the corresponding position of two arrays of the same size
Hamming loss: Hamming distance divided by number of array elements
Hamming loss is actually the Hamming distance divided by the total number of elements

1. 汉明距离是使用在数据传输差错控制编码里面的，汉明距离是一个概念，它表示两个（相同长度）字符串对应位置的
不同字符的数量，我们以d（x,y）表示两个字x,y之间的汉明距离。
2. 对两个字符串进行异或运算，并统计结果为1的个数，那么这个数就是汉明距离。
复制代码

2.2 Official

$\hat and$ To predict the result , $Y$ is thereal result, $y_{ij}$ For the i-th element in the j-th column , the maximum value of i is n , and the maximum value of j is m , in other words, $\hat y和y$ 均为 n*m大小的矩阵

$Hamming-loss(y, \hat{y}) = \frac{1}{n} \sum_{i=1}^{n} \frac{1}{m} \sum_{j=1}^{m}1*(\hat y_{ij} \neq y_{ij})$

2.3 应用场景

一般应用在多标签分类任务中
一般用作损失函数或者评价函数

[Hand tearing algorithm] [Keras] keras custom Hamming loss evaluation function