# -*- coding:utf-8 -*-
__author__ = 'snake'

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


def test01_notliner():
    # 使用numpy生成200个随机点
    x_data = np.linspace(-0.5, 0.5, 200).reshape((200,1))   # x_data维度设置为(200, 1)
    noise = np.random.normal(0, 0.02, x_data.shape)         # 生成干扰项
    y_data = np.square(x_data) + noise                      # 实际值

    # 定义两个placeholder
    x = tf.placeholder(tf.float32, [None, 1])
    y = tf.placeholder(tf.float32, [None, 1])

    # 定义神经网络中间层
    # x * W + b = pre_y
    Weights_L1 = tf.Variable(tf.random_normal([1, 10]))     # 10个神经元
    bases_L1 = tf.Variable(tf.zeros([1, 10]))
    Wx_plus_b_L1 = tf.matmul(x, Weights_L1) + bases_L1
    L1 = tf.nn.tanh(Wx_plus_b_L1)                           # 双曲正切函数作为激活函数

    # 定义神经网络输出层
    # 中间层的输出就是下一层的输入
    Weights_L2 = tf.Variable(tf.random_normal([10, 1]))     # 输出层1个神经元
    bases_L2 = tf.Variable(tf.zeros([1, 1]))                # 设置偏置值
    Wx_plus_b_L2 = tf.matmul(L1, Weights_L2) + bases_L2
    prediction = tf.nn.tanh(Wx_plus_b_L2)                   # 预测值
    
    # 代价函数
    loss = tf.reduce_mean(tf.square(y-prediction))
    # 使用梯度下降法最小化loss，学习速率（步长）为0.1
    train_step = tf.train.GradientDescentOptimizer(0.1).minimize(loss)
    
    with tf.Session() as sess:
        # 初始化变量
        sess.run(tf.global_variables_initializer())     # 初始化变量，在TensorFlow中用到Varibals就需要初始化

        # 循环2000次
        for _ in range(20000):
            # 运行梯度下降训练，训练数据为x_data, 实际值为y_data：
            # 实际上是监督学习：通过事先准备好的值来训练模型
            sess.run(train_step, feed_dict={x: x_data, y: y_data})

        # 获取预测值并画图
        prediction_value = sess.run(prediction, feed_dict={x: x_data, y: y_data})
        plt.figure()
        plt.scatter(x_data, y_data)
        plt.plot(x_data, prediction_value, 'r-', lw=5)
        plt.show()


if __name__ == "__main__":
    test01_notliner()