import tensorflow as tf
import numpy as np
import matplotlib.pyplot as plt
 
# 生成-0.5到0.5間均勻發(fā)布的200個點，將數(shù)據(jù)變?yōu)槎S，200行一列的數(shù)據(jù)
x_data = np.linspace(-0.5, 0.5, 200)[:, np.newaxis]
 
# 生成一些噪音數(shù)據(jù)
noise = np.random.normal(0, 0.02, x_data.shape)
 
# 定義y與x的關(guān)系
y_data = np.square(x_data) + noise
 
# 定義兩個占位符
x = tf.placeholder(tf.float32, [None, 1]) # 形狀為n行1列，同x_data的shape
y = tf.placeholder(tf.float32, [None, 1])
 
# 定義神經(jīng)網(wǎng)絡
 
# 定義中間層，因為每個x是一維，所以只需1個神經(jīng)元，定義中間層的連接神經(jīng)元是10
# 矩陣：[a, b]×[b, c] = [a, c] 
L1_weights = tf.Variable(tf.random_normal([1, 10])) 
L1_bias = tf.Variable(tf.zeros([1, 10]))
L1_weights_bias = tf.matmul(x, L1_weights) + L1_bias
L1 = tf.nn.tanh(L1_weights_bias)
 
# 定義輸出層，每個x只有一個神經(jīng)元
L2_weights = tf.Variable(tf.random_normal([10, 1]))
L2_bias = tf.Variable(tf.zeros([1, 1]))
L2_weights_bias = tf.matmul(L1, L2_weights) + L2_bias
L2 = tf.nn.tanh(L2_weights_bias)
 
# 定義損失函數(shù)
loss = tf.reduce_mean(tf.square(y - L2))
 
# 梯度下降最小化損失函數(shù)
optimizer = tf.train.GradientDescentOptimizer(0.1)
 
train_step = optimizer.minimize(loss)
 
# 全局變量初始化
init = tf.global_variables_initializer()
 
# 定義會話
with tf.Session() as sess:
 sess.run(init)
 for _ in range(2000):
  sess.run(train_step, feed_dict={x:x_data, y:y_data})
  
 # 獲取預測值
 predict = sess.run(L2, feed_dict={x:x_data})
 
 # 畫圖
 plt.figure()
 # 畫出散點
 plt.scatter(x_data, y_data)
 # 畫出擬合的曲線
 plt.plot(x_data, predict)
 
 plt.show()