tensorflow_DNN训练三角函数模型

标题利用简单的三层神经网络来训练sin函数和cos函数的模型，根据预测值画出相应的函数图像。tensorflow实现过程如下：

1. 配置超参数

class Config:
    def __init__(self):
        self.lr = 0.001
        self.epoches = 2000
        self.samples = 200	#训练样本数量
        self.hidden_units = 200  #隐藏层神经元个数

2. 获取训练样本

class Sample:
    def __init__(self, samples):
        self.xs = np.random.uniform(-np.pi, np.pi, [samples]) 	#随机正太分布生成-pi到pi的随机数作为样本x
        self.xs = sorted(self.xs)
        self.ys = np.sin(self.xs), np.cos(self.xs)  #样本目标值[2,-1]

3. 搭建神经网络

class Tensor:
    def __init__(self, config: Config):
        self.x = tf.placeholder(tf.float32, [None], name='x')  # [200]
        self.y = tf.placeholder(tf.float32, [2,None], name='y')	#[2,-1]
        x = tf.reshape(self.x, [-1, 1])  # 元素转化为张量
        x = tf.layers.dense(x, config.hidden_units, tf.nn.relu)
        self.y_predict = tf.layers.dense(x, 2)   #[-1,2]

        y = tf.transpose(self.y)     #使目标值的shape与预测值相同

        self.loss = tf.reduce_mean(tf.square(self.y_predict - y))  #计算loss
        self.lr = tf.placeholder(tf.float32, [], 'lr') 
        opt = tf.train.AdamOptimizer(self.lr)  #训练优化器
        self.train_op = opt.minimize(self.loss) 
        self.loss = tf.sqrt(self.loss)

4. 编写应用对象

class App:
    def __init__(self, config: Config):
        self.ts = Tensor(config)  #实例化张量网络
        self.config = config
        
        self.session = tf.Session()  #创建会话
        self.session.run(tf.global_variables_initializer())  #初始化变量

    def train(self):   #训练方法
        sample = Sample(200)
        for epoche in range(self.config.epoches):   
            _, loss = \
                self.session.run([self.ts.train_op, self.ts.loss],
                                 {self.ts.x: sample.xs, self.ts.y: sample.ys, self.ts.lr: self.config.lr})#喂入训练数据
            print(f"{epoche}:{loss}")
        return sample.xs, sample.ys  #返回样本数据

    def predict(self):  #预测方法
        sample = Sample(400)
        ys = self.session.run(self.ts.y_predict, {self.ts.x: sample.xs})   #喂入预测数据
        return sample.xs, ys  #返回预测数据

    def close(self):
        self.session.close()

5. 主方法

if __name__ == '__main__':
    app = App(Config())
    xs_raw, ys_raw = app.train()
    predict_x, predict_y = app.predict()
    plt.plot(predict_x, predict_y)
    plt.legend(['predict_sin','predict_cos'])
    plt.show()

6. 结果