TensorFlow seq2seq 回归模型

Question

我有一个简单的 seq2seq 模型来预测股票价格。我创建了一个 lstm 单元的编码器和解码器，它将预测接下来的 5 个时间步值。但是它会引发错误：

ValueError: Dimensions must be equal, but are 517 and 562 for 'rnn/while/rnn/multi_rnn_cell/cell_0/lstm_cell/MatMul_1' (op: 'MatMul') with input shapes: [10,517], [562,2048].

数据样本

               t1     t2     t3     t4    t5 ...
19/10/2018   0.005  0.100 -0.021 0.030 -0.025
20/10/2018   0.023  0.020  0.020 0.130  0.125
21/10/2018  -0.205  0.140 -0.011 0.020 -0.305

代码

import tensorflow as tf
import numpy as np

seq_len = 1
n_inputs = 50
n_outputs = 5
n_layers = 3
n_neurons = 512
batch_size = 10

g = tf.Graph()

with g.as_default():
  X = tf.placeholder(tf.float32,shape=(None,seq_len,n_inputs),name="X")
  y = tf.placeholder(tf.float32,shape=(None,seq_len,n_outputs),name="y")

  cells = tf.nn.rnn_cell.MultiRNNCell([ tf.nn.rnn_cell.LSTMCell(n_neurons) for _ in range(n_layers) ])

  init_state = cells.zero_state(batch_size, tf.float32)
  enc_outputs, enc_states = tf.nn.dynamic_rnn(cells, X,initial_state=init_state)

  dec_outputs,dec_states = tf.nn.dynamic_rnn(cells, y, initial_state=enc_states)

  loss = tf.reduce_mean(tf.square(dec_outputs - y))
  train_op = tf.train.AdamOptimizer(learning_rate=learning_rate).minimize(loss)

  init = tf.global_variables_initializer()

sess = tf.Session(graph=g)
sess.run(init)

欢迎任何帮助。

Answer 1

首先，我无法将您的问题标记为重复，因为它有赏金。您收到错误是因为您必须不要为第一层以及更深层重复使用相同的单元格。这是因为给它们的输入不同，这使得核矩阵不同。根据this 的帖子，这应该可以修复错误：

# Extra function is for readability. No problem to inline it.
def make_cell(lstm_size):
  return tf.nn.rnn_cell.BasicLSTMCell(lstm_size, state_is_tuple=True)

network = rnn_cell.MultiRNNCell([make_cell(num_units) for _ in range(num_layers)], 
                                state_is_tuple=True)

Here 在这个问题上有更多帮助。