在R中使用Keras 2.0构建股票预测LSTM模型时遇维度错误

Let's break down why you're getting that ValueError: Input 0 is incompatible with layer lstm_5: expected ndim=3, found ndim=2 error, and fix your code step by step.

Root Cause #

LSTM layers in Keras expect a 3D input tensor with the shape (number_of_samples, timesteps, number_of_features). Your current code has two key issues:

1. The dimensions of your x_train and x_test arrays are incorrect (you used 51 timesteps instead of 50).
2. The input_shape parameter in your LSTM layer was misconfigured.

Step 1: Fix Input Data Dimensions #

Looking at your code:

• You set seq_length = 50, so each training sample uses 50 days of data to predict the 51st day. That means each x_train entry should have 50 timesteps, not 51.
• Your original array conversion was hardcoding dimensions and not properly handling the list-to-array conversion. Instead, use unlist() to flatten the list first, then reshape it correctly.

Replace your array conversion code with this:

# Convert x_train list to 3D array: (samples, timesteps, features)
x_train <- array(unlist(x_train), dim = c(length(x_train), seq_length, 1))
# Convert y_train list to 2D array: (samples, 1)
y_train <- array(unlist(y_train), dim = c(length(y_train), 1))

# Do the same for test data
x_test <- array(unlist(x_test), dim = c(length(x_test), seq_length, 1))
y_test <- array(unlist(y_test), dim = c(length(y_test), 1))

Step 2: Correct the LSTM Layer Setup #

The input_shape parameter for an LSTM layer refers to the shape of a single sample (omitting the number of samples, which Keras handles automatically). Since each of your samples has 50 timesteps and 1 feature (closing price), set input_shape = c(seq_length, 1).

Here's the fixed model code:

library(keras)
model <- keras_model_sequential()

# First LSTM layer: 50 units, return sequences for stacking another LSTM (if needed)
model %>% 
  layer_lstm(units = 50, return_sequences = TRUE, input_shape = c(seq_length, 1)) %>%
  # Second LSTM layer (optional but common for time series)
  layer_lstm(units = 50) %>%
  # Output layer: predicts a single value (next day's normalized price)
  layer_dense(units = 1)

# Compile the model with appropriate loss and optimizer
model %>% compile(
  loss = 'mean_squared_error',
  optimizer = optimizer_adam(learning_rate = 0.001)
)

Key Notes #

• If you want the model to accept variable-length timesteps (not just 50), you can use input_shape = c(NULL, 1) instead. Just make sure all samples in a batch have the same timestep length during training.
• The units parameter replaces the old output_dim from Keras 1.x—this sets the number of hidden units in the LSTM layer (50 is a reasonable starting point for this dataset).

Why Your Previous Input Shapes Failed #

• input_shape = c(1L): This tells Keras to expect 2D input (samples, 1), which doesn't match the 3D input LSTMs require.
• input_shape = c(NULL, 1L): While this is valid for variable timesteps, your input data had the wrong timestep count (51 instead of 50), so Keras still threw a dimension mismatch error.

内容的提问来源于stack exchange，提问作者Mislav