tensorflow1.1/利用rnn回归分析

环境：tensorflow1.1,python3,matplotlib2.02

#coding:utf-8from tensorflow.contrib import rnnimport tensorflow as tfimport numpy as npimport matplotlib.pyplot as plttime_step = 1input_size = 1n_hidden_units = 32batch_size = 128learning_rate = 0.01#构造数据x = np.linspace(0,np.pi*5,1000,dtype=np.float32)noise = np.random.normal(0,0.1,size=x.shape)y = np.sin(x) + noisexs = tf.placeholder(tf.float32,[None,time_step,input_size])#输入和输出形状相同ys = tf.placeholder(tf.float32,[None,time_step,input_size])#定义weights和biasesweights = {    'in':tf.Variable(tf.random_normal([input_size,n_hidden_units])),    'out':tf.Variable(tf.random_normal([n_hidden_units,input_size]))}biases = {    'in':tf.Variable(tf.constant(0.1,shape=[n_hidden_units])),    'out':tf.Variable(tf.constant(0.1,shape=[input_size]))}#RNN输出二维数据def RNN(x,weights,biases):    #输入数据形状(batch_size,time_step,input_size)------>RNN输入形状(batch_size,input_size)    x = tf.unstack(x,time_step,1) #将time_step维度去除    lstm_cell = rnn.BasicLSTMCell(n_hidden_units,forget_bias=1.0)    #返回outputs,主线state,分线state    outputs,states = rnn.static_rnn(lstm_cell,x,dtype=tf.float32)    return tf.matmul(outputs[-1],weights['out'])+biases['out']output = RNN(xs,weights,biases)#将数据转为三维output_reshape = tf.reshape(output,[-1,time_step,input_size])cost = tf.losses.mean_squared_error(labels=ys,predictions=output_reshape)train = tf.train.AdamOptimizer(learning_rate).minimize(cost)init = tf.global_variables_initializer()with tf.Session() as sess:    sess.run(init)    a = plt.figure(1,figsize=(6,5))    #开始交互模式    plt.ion()    for step in range(5000):        k = 0        while k < x.shape[0]:            batch_x = x[k:k+batch_size]            batch_y = y[k:k+batch_size]            #将输入转为三维数据            batch_x = batch_x.reshape(-1,time_step,input_size)            batch_y = batch_y.reshape(-1,time_step,input_size)            k = k+batch_size            _,c = sess.run([train,cost],feed_dict={xs:batch_x,ys:batch_y})        outputs = sess.run(output,feed_dict={xs:x.reshape(-1,time_step,input_size),ys:y.reshape(-1,time_step,input_size)})        if step % 500 ==0:            a.clear()            plt.plot(x,y,'r-',label='real data')            plt.legend(loc='best')            plt.plot(x,outputs,'b-',label='regression data')            plt.legend(loc='best')            plt.grid(True)            plt.show()            plt.pause(0.1)    plt.ioff()

结果：

随着训练次数的增加,拟合的效果越好