我在训练具有多列输入输出的LSTM模型时遇到了一个问题。我的代码如下：
time_step = 60
#Create a data structure with n-time steps
X = []
y = []
for i in range(time_step + 1, len(training_set_scaled)):
X.append(training_set_scaled[i-time_step-1:i-1, 0:len(training_set.columns)]) #take all columns into the set
y.append(training_set_scaled[i, 0:len(training_set.columns)]) #take all columns into the set
X_train_arr, y_train_arr = np.array(X), np.array(y)
print(X_train_arr.shape) #(2494, 60, 5)
print(y_train_arr.shape) #(2494, 5)
#Split data
X_train_splitted = X_train_arr[:split]
y_train_splitted = y_train_arr[:split]
X_test_splitted = X_train_arr[split:]
y_test_splitted = y_train_arr[split:]
#Initialize the RNN
model = Sequential()
#Add the LSTM layers and some dropout regularization
model.add(LSTM(units= 50, activation = 'relu', return_sequences = True, input_shape = (X_train_arr.shape[1], X_train_arr.shape[2]))) #time_step/columns
model.add(Dropout(0.2))
model.add(LSTM(units= 40, activation = 'relu', return_sequences = True))
model.add(Dropout(0.2))
model.add(LSTM(units= 80, activation = 'relu', return_sequences = True))
model.add(Dropout(0.2))
#Add the output layer.
model.add(Dense(units = 1))
#Compile the RNN
model.compile(optimizer='adam', loss = 'mean_squared_error')
#Fit to the training set
model.fit(X_train_splitted, y_train_splitted, epochs=3, batch_size=32)
其思想是从i后退60步并在i:中具有5列目标来训练模型
for i in range(time_step + 1, len(training_set_scaled)):
X.append(training_set_scaled[i-time_step-1:i-1, 0:len(training_set.columns)]) #take all columns into the set
y.append(training_set_scaled[i, 0:len(training_set.columns)]) #take all columns into the set
所以我的x列(馈送(和y列(目标(是：
X_train_arr, y_train_arr = np.array(X), np.array(y)
print(X_train_arr.shape) #(2494, 60, 5)
print(y_train_arr.shape) #(2494, 5)
不幸的是，在拟合模型时：
model.fit(X_train_splitted, y_train_splitted, epochs=3, batch_size=32)
我得到一个错误：
维度必须相等，但对于"｛｛node"为60和5mean_squared_error/SquaredDifference｝｝=平方差[T=DT_FLOAT](mean_squared_error/remove_squeezable_dimensions/Squeeze，IteratorGetNext:1('，具有输入形状：[？，60]，[？，5]。
我知道X_train_arr和y_train_arr需要相同。但当用下面的案例进行测试时，一切都很好：
X_train_arr, y_train_arr = np.array(X), np.array(y)
print(X_train_arr.shape) #(2494, 60, 5)
print(y_train_arr.shape) #(2494, 1)
拥有print(y_train_arr.shape) #(2494, 5)的想法是能够预测未来的n步，其中每次预测迭代都会生成具有5列值的新的整行数据。
好吧，完成本教程后，我明白了应该做什么。以下是带有注释的最终代码：
#Variables
future_prediction = 30
time_step = 60 #learning step
split_percent = 0.80 #train/test data split percent (80%)
split = int(split_percent*len(training_set_scaled)) #split percent multiplying by data rows
#Create a data structure with n-time steps
X = []
y = []
for i in range(time_step + 1, len(training_set_scaled)):
X.append(training_set_scaled[i-time_step-1:i-1, 0:len(training_set.columns)]) #take all columns into the set, including time_step legth
y.append(training_set_scaled[i, 0:len(training_set.columns)]) #take all columns into the set
X_train_arr, y_train_arr = np.array(X), np.array(y) #must be numpy array for TF inputs
print(X_train_arr.shape) #(2494, 60, 5) <-- train data, having now 2494 rows, with 60 time steps, each row has 5 features (MANY)
print(y_train_arr.shape) #(2494, 5) <-- target data, having now 2494 rows, with 1 time step, but 5 features (TO MANY)
#Split data
X_train_splitted = X_train_arr[:split] #(80%) model train input data
y_train_splitted = y_train_arr[:split] #80%) model train target data
X_test_splitted = X_train_arr[split:] #(20%) test prediction input data
y_test_splitted = y_train_arr[split:] #(20%) test prediction compare data
#Reshaping to rows/time_step/columns
X_train_splitted = np.reshape(X_train_splitted, (X_train_splitted.shape[0], X_train_splitted.shape[1], X_train_splitted.shape[2])) #(samples, time-steps, features), by default should be already
y_train_splitted = np.reshape(y_train_splitted, (y_train_splitted.shape[0], 1, y_train_splitted.shape[1])) #(samples, time-steps, features)
X_test_splitted = np.reshape(X_test_splitted, (X_test_splitted.shape[0], X_test_splitted.shape[1], X_test_splitted.shape[2])) #(samples, time-steps, features), by default should be already
y_test_splitted = np.reshape(y_test_splitted, (y_test_splitted.shape[0], 1, y_test_splitted.shape[1])) #(samples, time-steps, features)
print(X_train_arr.shape) #(2494, 60, 5)
print(y_train_arr.shape) #(2494, 1, 5)
print(X_test_splitted.shape) #(450, 60, 5)
print(y_test_splitted.shape) #(450, 1, 5)
#Initialize the RNN
model = Sequential()
#Add Bidirectional LSTM, has better performance than stacked LSTM
model = Sequential()
model.add(Bidirectional(LSTM(100, activation='relu', input_shape = (X_train_splitted.shape[1], X_train_splitted.shape[2])))) #input_shape will be (2494-size, 60-shape[1], 5-shape[2])
model.add(RepeatVector(5)) #for 5 column of features in output, in other cases used for time_step in output
model.add(Bidirectional(LSTM(100, activation='relu', return_sequences=True)))
model.add(TimeDistributed(Dense(1)))
#Compile the RNN
model.compile(optimizer='adam', loss = 'mean_squared_error')
#Fit to the training set
model.fit(X_train_splitted, y_train_splitted, epochs=3, batch_size=32, validation_split=0.2, verbose=1)
#Test results
y_pred = model.predict(X_test_splitted, verbose=1)
print(y_pred.shape) #(450, 5, 1) - need to be reshaped for (450, 1, 5)
#Reshaping data for inverse transforming
y_test_splitted = np.reshape(y_test_splitted, (y_test_splitted.shape[0], 5)) #reshaping for (450, 1, 5)
y_pred = np.reshape(y_pred, (y_pred.shape[0], 5)) #reshaping for (450, 1, 5)
#Reversing transform to get proper data values
y_test_splitted = scaler.inverse_transform(y_test_splitted)
y_pred = scaler.inverse_transform(y_pred)
#Plot data
plt.figure(figsize=(14,5))
plt.plot(y_test_splitted[-time_step:, 3], label = "Real values") #I am interested only with display of column index 3
plt.plot(y_pred[-time_step:, 3], label = 'Predicted values') # #I am interested only with display of column index 3
plt.title('Prediction test')
plt.xlabel('Time')
plt.ylabel('Column index 3')
plt.legend()
plt.show()
#todo: future prediction