tensorflow中交叉熵损失函数预测酸奶价格

tensorflow中交叉熵损失函数

引用API：tensorflow.kreas.losses.categorical_crossentropy

则，tf.losses.categorical_crossentropy(y_, y) #y_为真实值，y为预测值

 交叉熵是说明两个概率分布之间的距离，距离越小，说明二者分布越接近

例如：已知真实值y_(1, 1),预测值y1(0.8, 0.6),预测值y2(0.2, 0.4)。这两个预测值哪一个距离真实值最近。

可得：H1((1, 1), (0.8, 0.5)) = -(1 * ln(0.8) + 1 * ln(0.5)) ~0.69

           H2((1, 2), (0.2, 0.4)) = -(1 * ln(0.2) + 1 * ln(0.4)) ~0.92

因为H1<H2，所以H1更接近真实值

 tensorflow中交叉熵损失函数预测酸奶价格

import tensorflow as tf
import numpy as np

seed = 23456 #定义种子数值

rdm = np.random.RandomState(seed) #生成[0, 1)之间的小数,由于固定seed数值，每次生成的数值都一样

x = rdm.rand(32, 2) #生成32行2列的输入特征x

np.set_printoptions(threshold=6) #概略显示

print("32行2列的特征xs为：", x)  # 打印输入特征，概略显示5行

y_ = [[x1 + x2 + (rdm.rand()  / 10.0 - 0.05)] for (x1, x2) in x] #从x中取出x1,x2相加，并且加上-0.05到+0.05随机噪声，得到标准答案y_

x = tf.cast(x, dtype = tf.float32) #转换x数据类型

print("转换数据类型后x:", x) #打印转换书籍类型后的x

w1 = tf.Variable(tf.random.normal([2, 1], stddev = 1, seed = 1)) #初始化参数w1，生成两行一列的可训练参数

print("w1为：", w1)

epoch = 15000 #迭代次数为15000次

lr = 0.002 #学习率为0.002

for epoch in range(epoch):
    with tf.GradientTape() as tape: #用with上下文管理器调用GradientTape()方法
        y = tf.matmul(x, w1) #求前向传播计算结果y
        # loss_mse = tf.reduce_mean(tf.square(y_ - y)) #求均方误差loss_mean
        loss_mse = tf.losses.categorical_crossentropy(y_, y)  # 求交叉熵loss_mean
# print("y为：", y)
# print("loss_mean为：", loss_mse)
    grads = tape.gradient(loss_mse, w1) #损失函数对待训练参数w1求偏导
    w1.assign_sub(lr * grads) #更新参数w1

    if epoch % 500 == 0: #每迭代500轮
        print("After %d training steps, w1 is " % epoch) #每迭代500轮打印当前w1参数
        print(w1.numpy(), "\n") #打印当前w1参数

print("Final w1 is :", w1.numpy()) #打印最终w1

结果为：

E:\Anaconda3\envs\TF2\python.exe C:/Users/Administrator/PycharmProjects/untitled8/酸奶价格预测.py
32行2列的特征xs为： [[0.32180029 0.32730047]
 [0.92742231 0.31169778]
 [0.16195411 0.36407808]
 ...
 [0.83554249 0.59131388]
 [0.2829476  0.05663651]
 [0.2916721  0.33175172]]
转换数据类型后x: tf.Tensor(
[[0.3218003  0.32730046]
 [0.9274223  0.31169778]
 [0.1619541  0.36407807]
 ...
 [0.8355425  0.5913139 ]
 [0.2829476  0.05663651]
 [0.29167208 0.33175173]], shape=(32, 2), dtype=float32)
w1为： <tf.Variable 'Variable:0' shape=(2, 1) dtype=float32, numpy=
array([[-0.8113182],
       [ 1.4845988]], dtype=float32)>
After 0 training steps, w1 is 
[[-0.8113182]
 [ 1.4845988]] 

After 500 training steps, w1 is 
[[-0.8113182]
 [ 1.4845988]] 

After 1000 training steps, w1 is 
[[-0.8113182]
 [ 1.4845988]] 

After 1500 training steps, w1 is 
[[-0.8113182]
 [ 1.4845988]] 

After 2000 training steps, w1 is 
[[-0.8113182]
 [ 1.4845988]] 

After 2500 training steps, w1 is 
[[-0.8113182]
 [ 1.4845988]] 

After 3000 training steps, w1 is 
[[-0.8113182]
 [ 1.4845988]] 

After 3500 training steps, w1 is 
[[-0.8113182]
 [ 1.4845988]] 

After 4000 training steps, w1 is 
[[-0.8113182]
 [ 1.4845988]] 

After 4500 training steps, w1 is 
[[-0.8113182]
 [ 1.4845988]] 

After 5000 training steps, w1 is 
[[-0.8113182]
 [ 1.4845988]] 

After 5500 training steps, w1 is 
[[-0.8113182]
 [ 1.4845988]] 

After 6000 training steps, w1 is 
[[-0.8113182]
 [ 1.4845988]] 

After 6500 training steps, w1 is 
[[-0.8113182]
 [ 1.4845988]] 

After 7000 training steps, w1 is 
[[-0.8113182]
 [ 1.4845988]] 

After 7500 training steps, w1 is 
[[-0.8113182]
 [ 1.4845988]] 

After 8000 training steps, w1 is 
[[-0.8113182]
 [ 1.4845988]] 

After 8500 training steps, w1 is 
[[-0.8113182]
 [ 1.4845988]] 

After 9000 training steps, w1 is 
[[-0.8113182]
 [ 1.4845988]] 

After 9500 training steps, w1 is 
[[-0.8113182]
 [ 1.4845988]] 

After 10000 training steps, w1 is 
[[-0.8113182]
 [ 1.4845988]] 

After 10500 training steps, w1 is 
[[-0.8113182]
 [ 1.4845988]] 

After 11000 training steps, w1 is 
[[-0.8113182]
 [ 1.4845988]] 

After 11500 training steps, w1 is 
[[-0.8113182]
 [ 1.4845988]] 

After 12000 training steps, w1 is 
[[-0.8113182]
 [ 1.4845988]] 

After 12500 training steps, w1 is 
[[-0.8113182]
 [ 1.4845988]] 

After 13000 training steps, w1 is 
[[-0.8113182]
 [ 1.4845988]] 

After 13500 training steps, w1 is 
[[-0.8113182]
 [ 1.4845988]] 

After 14000 training steps, w1 is 
[[-0.8113182]
 [ 1.4845988]] 

After 14500 training steps, w1 is 
[[-0.8113182]
 [ 1.4845988]] 

Final w1 is : [[-0.8113182]
 [ 1.4845988]]

Process finished with exit code 0