tensorflow2.0(2)--线性回归

最新推荐文章于 2022-12-15 02:00:00 发布

CRwake

最新推荐文章于 2022-12-15 02:00:00 发布

阅读量279

点赞数

本文链接：https://blog.csdn.net/CRwake/article/details/105892902

版权

线性回归

import tensorflow as tf
import os
import matplotlib.pyplot as plt
import numpy as np

os.environ['TF_CPP_MIN_LOG_LEVEL']='2'		

data=3*np.arange(1000)+np.random.rand(1000)*100		#制造数据y=3x

# plt.plot(data)
# plt.show()

train_w=tf.constant(0.0)			#注意要使用0.0，dtype:float32
train_b=tf.constant(0.0)

true_y=tf.convert_to_tensor(data)	
true_y=tf.cast(true_y,dtype=tf.float32)	#能与后面进行运算
true_x=tf.range(1000.0)					#注意使用.0

train_dataset = tf.data.Dataset.from_tensor_slices((true_x, true_y))	
train_dataset = train_dataset.batch(200)

for i in range(10):
    for step, (true_x,true_y) in enumerate(train_dataset):
        with tf.GradientTape() as tape:
            tape.watch([train_b,train_w])		#如果是tf.Variable类型就不用
            y=train_w*true_x+train_b
            loss=tf.reduce_sum((y-true_y)**2)/true_x.shape[0]	#等于loss=tf.reduce_mean((y-true_y)**2)
        grads=tape.gradient(loss,[train_b,train_w])	#自动求导
        print(grads[1].numpy())						#可观察大小，确定学习率

        train_b-=(grads[0]*1e-7)					#学习率太大，会导致梯度正负变化，且loss会趋于inf
        train_w-=(grads[1]*1e-7)

        if step %2 ==0:
            print('i=%i,w=%f,b=%f,loss=%f' %(step,train_w.numpy(),train_b.numpy(),loss.numpy()))

print(train_w.numpy(),train_b.numpy())