Question

我正在尝试使用GRU基于深度学习构建时间序列预测。我设法得到一个预测。供参考，如果有人想复制我的过程，这里是我的数据-https://drive.google.com/file/d/1nTQjkMuYFVaYFP4mTtjTZkEXCLwjmGip/view?usp=sharing

这是我的数据。我正在尝试预测“总计。总和”功能-

str(df)

'data.frame':   268 obs. of  5 variables:
 $ Total.Sum: num  0.0159 0.0233 0.0234 0.024 0.0157 ...
 $ weekday  : num  0.25 0.5 0.75 1 0 0.25 0.5 0.75 1 0 ...
 $ holiday  : num  0 0 0 0 0 0 0 0 0 0 ...
 $ date     : num  0.978 0.978 0.978 0.978 0.978 ...
 $ month    : num  0.0201 0.0201 0.0201 0.0201 0.0201 ...

我有256个训练示例。预处理模型数据-

# Splitting into training and testing set
df <- as.matrix(df)
x_train <- array_reshape(df[1:256,], dim = c(256,5,1)) # shifting testing set by 1 timestep
y_train <- df[,1] # first column is our target variable
y_train <- array_reshape(df, dim = c(256, 1))

这是我的GRU模型-

model <- keras_model_sequential()
model %>% 
  layer_gru(units = 32, batch_input_shape = c(256, 5, 1), return_sequences = T, stateful = T, activation = "relu") %>%
  layer_gru(units = 64, return_sequences = T, stateful = T, activation = "relu") %>%
  layer_gru(units = 128, return_sequences = T, stateful = T, activation = "relu") %>%
  layer_gru(units = 32, return_sequences = F, stateful = T, activation = "relu") %>%
  layer_dense(units = 1) %>%
  layer_activation(activation = "linear")

model %>% compile(loss = "mse", optimizer = "adam")

最后，这是我的训练循环-

for (i in 1:1000) {
  model %>% fit(x_train, y_train, batch_size = 256, shuffle = FALSE, epochs = 1)
  model %>% reset_states()
}

我的问题是，如何获得模型的预测以获取长期预测？我一直在尝试从这篇Understanding Keras LSTMs帖子中了解这是如何实现的，但发现很难理解，尤其是在尝试将python转换为R时。据我所知，我们需要预测下一步在Y中，然后以某种方式将其循环回到我们的模型中...

这就是我走的路-

X <- x_train[-dim(totalSequences)[1],,] #the entire known sequence, except the last step
X <- array_reshape(X, dim = c(255, 5, 1))

Y <- x_train[dim(x_train)[1],,] #one step ahead of X
Y <- array_reshape(Y, dim = c(1, 5))

#loop for resetting states at the start/end of the sequences:
for (i in 1:60){
  model %>% reset_states()
  model %>% fit(X, Y, batch_size = 256, shuffle = FALSE, epochs = 10)
}

这是我遇到无法解决的形状不匹配错误的地方。任何帮助都将不胜感激！

Error in py_call_impl(callable, dots$args, dots$keywords) : 
  ValueError: Error when checking target: expected activation_14 to have shape (1,) but got array with shape (5,)

使用多元GRU（Keras，R）预测长期预测

0 个答案: