如何正确塑造 RNN 的时间序列数据？答案

【问题标题】：How to correctly shape time-series data for RNN?如何正确塑造 RNN 的时间序列数据？
【发布时间】：2019-07-31 19:54:20
【问题描述】：

我已经开始使用 Python 中的 TensorFlow 进行一个简单的项目，以使用循环网络预测股票市场价格。到目前为止，这是我的代码：

n_steps = 30
n_inputs = 1
n_neurons = 100
n_outputs = 1

X = tf.placeholder(tf.float32, [1, n_steps, n_inputs])
y = tf.placeholder(tf.float32, [None, n_steps, n_inputs])
cell = tf.contrib.rnn.OutputProjectionWrapper(
    tf.contrib.rnn.BasicRNNCell(num_units=n_neurons, activation=tf.nn.relu),
    output_size = n_outputs
)
outputs, states = tf.nn.dynamic_rnn(cell, X, dtype=tf.float32)

learning_rate = 0.001

loss = tf.reduce_mean(tf.square(outputs - y))
optimizer = tf.train.AdamOptimizer(learning_rate=learning_rate)
training_op = optimizer.minimize(loss)

init = tf.global_variables_initializer()
n_iterations = numStocks
batch_size = 1

def priceArrayToRNNFormat(priceArray):
    list = []
    print(priceArray)
    for price in priceArray:
        list.append(price)
    return np.array(list)

with tf.Session() as sess:
    init.run()
    for iteration in range(n_iterations):
        dataOrig = [allStocksDict[list(allStocksDict.keys())[iteration]]]
        data = priceArrayToRNNFormat(dataOrig)
        print(data)
        X_batch = data
        y_batch = data
        sess.run(training_op, feed_dict={X: X_batch, y: y_batch})
        if iteration % 100 == 0:
            mse = loss.eval(feed_dict={X: X_batch, y: y_batch})
            print(iteration, "\tMSE", mse)

作为参考，allStocksDict 只是一个字典，其中每个键都是一个股票代码，值是一个 30 元素的数组，其中包含一段时间内的价格。运行代码时，我得到以下输出：

[['14.9400', '15.0000', '14.8800', '14.6900', '14.6300', '15.0000', '14.9400', '15.1300', '15.5600', '15.3100', '15.3800', '14.6900', '15.0000', '15.1300', '14.6300', '14.0600', '14.1300', '14.9400', '14.4400', '13.6300', '13.0000', '12.3800', '12.5000', '12.6300', '13.0000', '12.6900', '13.1300', '13.1900', '13.0600', '12.9400']]
[['14.9400' '15.0000' '14.8800' '14.6900' '14.6300' '15.0000' '14.9400'
  '15.1300' '15.5600' '15.3100' '15.3800' '14.6900' '15.0000' '15.1300'
  '14.6300' '14.0600' '14.1300' '14.9400' '14.4400' '13.6300' '13.0000'
  '12.3800' '12.5000' '12.6300' '13.0000' '12.6900' '13.1300' '13.1900'
  '13.0600' '12.9400']]
Traceback (most recent call last):
  File "/home/john/Python/StockProject/monthlyRnn1.py", line 127, in <module>
    sess.run(training_op, feed_dict={X: X_batch, y: y_batch})
  File "/home/john/.local/lib/python3.6/site-packages/tensorflow/python/client/session.py", line 929, in run
    run_metadata_ptr)
  File "/home/john/.local/lib/python3.6/site-packages/tensorflow/python/client/session.py", line 1128, in _run
    str(subfeed_t.get_shape())))
ValueError: Cannot feed value of shape (1, 30) for Tensor 'Placeholder:0', which has shape '(1, 30, 1)'

我尝试自行提供列表而不将其转换为数组，并且在将其转换为数组之前未将数组转换为向量，尽管此错误仍然存在。非常感谢您对此的帮助。

【问题讨论】：

标签： python arrays numpy tensorflow neural-network

【解决方案1】：

一种可能的解决方案是

def priceArrayToRNNFormat(priceArray):
    #list = []
    #print(priceArray)
    #for price in priceArray:
    #    list.append(price)
    #return np.array(list)
    return np.reshape(np.asarray(priceArray, dtype=np.float32), (1, n_steps, n_inputs))

嵌套列表也是可以接受的，另一种选择是转置 priceArray 并将其再次包装成一个小批量列表。
但前一个选项 np.reshape() 简单快捷。

【讨论】：