TensorFlow 使用多项式进行sin函数的拟合操作

定义一个主函数,对sin函数进行显示：

if __name__ == "__main__":samples = 200xs = np.random.uniform(-np.pi, np.pi, [samples])xs = sorted(xs)ys = sorted(ys)# 使用matplotlib进行图像的显示plt.plot(xs, ys)plt.show()

在主方法中调用SinApp类

app = SinApp(Config())
with app:xs_train, ys_train = app.train()xs_predict, ys_predict = app.predict()
plt.plot(xs_train,ys_train)
plt.plot(xs_predict, ys_predict)
plt.show()

将用到的参数放在Config的类中

class Config:def __init__(self):self.save_path = './model_sin/sin'self.lr = 0.001self.epoches = 2000self.batch_size = 200# 定义隐藏层的数量self.hidden_units = 200

定义一个张量类：Tensors

class Tensors:def __init__(slef, config):self.x = tf.placeholder(tf.float32, [None], 'x')self.y = tf.placeholder(tf.float32, [None], 'y')x = tf.reshape(slef.x, [-1, 1])x = tf.layer.dense(x, config.hidden_units, tf.nn.relu)y = tf.layer.dense(x, 1)self.y_predict = tf.reshape(y, [-1])self.loss = tf.reduce_mean(tf.square(tf.y_predict - self.y))  # 使用方差损失self.lr = tf.placeholder(tf.float32, [], 'lr')  # 定义学习步长（可以定义成动态的）opt = tf.train.AdaOptimizer(self.lr)  # 定义优化器self.train_op = opt.minimize(self.loss)self.loss = tf.sqrt(self.loss)  # 取sinx的平方根（求平方差）打印的会更合理的，减少误差

定义一个样本类：Sample（实际上大部分工作都是在处理样本的）

class Sample:def __init__(self, samples):self.xs = np.random.uniform(-np.pi, np.pi, [samples])  # 可以自己定义samplesself.xs = sorted(self.xs)self.ys = np.sin(self.xs)@propertydef num_examples(self):return len(self.xs)

定义SinApp类

class SinApp:def __init__(self, config):self.ts = Tensors(config)self.session = tf.Session()self.saver = tf.train.Saver()try:self.saver.restore(self.session, config.save_path)except:self.session.run(tf.global_variables_initializer())def train(self):sample = Sample(self.config.samples)cfg = self.configts = self.tsfor _ in range(cfg.epoches):_ , loss = self.session.run([ts.train_op, ts.loss], {ts.x: sample.xs, ts.y:sample.ys, ts.lr:cfg.lr})self.save()return samples.xs, samples.ys  # 通过训练得到的xs和ysdef	save(self):self.saver.save(self.session, self.config.save_path)print('save model into', self.save_path)def predict(self):sample = Sample(400)  # 样本数量400个，不是很重要就不写在config中了ys = self.session.run(self.ts.y_predict, {self.ts.x: samples.xs})  # 预测的值弄出来，按照顺序算出ys的值了。return sample.xs, ys  # 返回xs是400个样本点，ys是对应的正弦值def close(self):self.session.close()def __enter__(self):return selfdef __exit__(self):self.close()

运行结果为：

局部放大图：