1. 预处理

url = './铁路客运量.csv'

ass_data = requests.get(url).content #打开文件

df = pd.read_csv(io.StringIO(ass_data.decode('utf-8')))  # python2使用StringIO.StringIO

data = np.array(df['铁路客运量_当期值(万人)']) 

# normalize

normalized_data = (data - np.mean(data)) / np.std(data)

seq_size = 3

train_x, train_y = [], []

for i in range(len(normalized_data) - seq_size - 1): # 解析

 train_x.append(np.expand_dims(normalized_data[i: i + seq_size], axis=1).tolist())

    train_y.append(normalized_data[i + 1: i + seq_size + 1].tolist())

input_dim = 1

X = tf.placeholder(tf.float32, [None, seq_size, input_dim])

Y = tf.placeholder(tf.float32, [None, seq_size])

2. 模型构建

def ass_rnn(hidden_layer_size=6): #  regression

    W = tf.Variable(tf.random_normal([hidden_layer_size, 1]), name='W')

    b = tf.Variable(tf.random_normal([1]), name='b')

    cell = tf.nn.rnn_cell.BasicLSTMCell(hidden_layer_size)# 6

    outputs, states = tf.nn.dynamic_rnn(cell, X, dtype=tf.float32)

    W_repeated = tf.tile(tf.expand_dims(W, 0), [tf.shape(X)[0], 1, 1])

    out = tf.batch_matmul(outputs, W_repeated) + b

    out = tf.squeeze(out)

    return out

3. 训练模型

def train_rnn():

    out = ass_rnn()

    loss = tf.reduce_mean(tf.square(out - Y))

    train_op = tf.train.AdamOptimizer(learning_rate=0.003).minimize(loss)

    saver = tf.train.Saver(tf.all_variables())  # tf.global_variables()  == tf 0.12

    with tf.Session() as sess:

        tf.get_variable_scope().reuse_variables()  # old API

        sess.run(tf.initialize_all_variables())  # tf.global_variables_initializer() ==tf  0.12

        for step in range(10000):

            _, loss_ = sess.run([train_op, loss], feed_dict={X: train_x, Y: train_y})

            if step % 10 == 0:

                # 用测试数据评估loss

                print(step, loss_)

        print("保存模型: ", saver.save(sess, 'ass.model'))

4. 预测

def prediction():

    out = ass_rnn()

    saver = tf.train.Saver(tf.all_variables())  # new API  tf.global_variables()

    with tf.Session() as sess:

        tf.get_variable_scope().reuse_variables()  # old API

        saver.restore(sess, './ass.model')

        prev_seq = train_x[-1]

        predict = []

        for i in range(12):

            next_seq = sess.run(out, feed_dict={X: [prev_seq]})

            predict.append(next_seq[-1])

            prev_seq = np.vstack((prev_seq[1:], next_seq[-1]))

        plt.figure()

        plt.plot(list(range(len(normalized_data))), normalized_data, color='b')

        plt.plot(list(range(len(normalized_data), len(normalized_data) + len(predict))), predict, color='r')

        plt.show()

# train_rnn()  #

prediction()