首先构建一个线性的点状图
import warnings
warnings.filterwarnings('ignore')
import numpy as np
import matplotlib.pyplot as plt
%matplotlib inline
from sklearn.linear_model import LinearRegression
import tensorflow as tf
X = np.linspace(2,12,50).reshape(-1,1)
w = np.random.randint(1,6,size = 1)[0]
b = np.random.randint(-5,5,size = 1)[0]
y = X*w + b + np.random.randn(50,1)*0.7
plt.scatter(X,y)
查看X,y的类型
print(X.shape,y.shape)
(50, 1) (50, 1)
使用线性回归查看预测的系数为了对比TF
linear = LinearRegression()
linear.fit(X,y)
print(linear.coef_,linear.intercept_)
[[3.05044116]] [-1.33814071]
print(w,b) #最后测试对比
3 -1
Tensorflow完成线性回归
1、定义占位符、变量
# 线性回归理论基础是最小二乘法
X_train = tf.placeholder(dtype=tf.float32,shape = [50,1],name = 'data')
y_train = tf.placeholder(dtype=tf.float32,shape = [50,1],name = 'target')
w_ = tf.Variable(initial_value=tf.random_normal(shape = [1,1]),name = 'weight')
b_ = tf.Variable(initial_value=tf.random_normal(shape = [1]),name = 'bias')
2、构造方程(线性方程,矩阵乘法)
# 构建方程 f(x) = Xw + b
# 构建的方程,就是预测的结果
y_pred = tf.matmul(X_train,w_) + b_
# shape = (50,1)
y_pred
<tf.Tensor ‘add:0’ shape=(50, 1) dtype=float32>
3、最小二乘法(平均最小二乘法)
# 二乘法(y_pred - y_train)**2 返回的结果是列表,没有办法比较大小
# 平均最小二乘法,数值,mean
# 平均:每一个样本都考虑进去了
cost = tf.reduce_mean(tf.pow(y_pred - y_train,2))
cost
<tf.Tensor ‘Mean_1:0’ shape=() dtype=float32>
4、梯度下降(tf,提供了方法)
# 优化,cost损失函数,越小越好
opt = tf.train.GradientDescentOptimizer(0.01).minimize(cost)
opt
<tf.Operation ‘GradientDescent_1’ type=NoOp>
5、会话进行训练(for循环),sess.run(),占位符(赋值)
with tf.Session() as sess:
# 变量,初始化
sess.run(tf.global_variables_initializer())
for i in range(1000):
opt_,cost_ = sess.run([opt,cost],feed_dict = {y_train:y,X_train:X})
if i %50 == 0:
print('执行次数是:%d。损失函数值是:%0.4f'%(i+1,cost_))
# for循环结束,训练结束了
# 获取斜率和截距
W,B = sess.run([w_,b_])
print('经过100次训练,TensorFlow返回线性方程的斜率是:%0.3f。截距是:%0.3f'%(W,B))
执行次数是:1。损失函数值是:581.9765
执行次数是:51。损失函数值是:1.3099
执行次数是:101。损失函数值是:1.0826
。。。
执行次数是:951。损失函数值是:0.4290
经过100次训练,TensorFlow返回线性方程的斜率是:3.033。截距是:-1.194
6、可视化
plt.scatter(X,y)
x = np.linspace(0,14,100)
plt.plot(x,W[0]*x + B,color = 'green')