I am facing a problem with training an LSTM model with multicolumn input output. My code is below:
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)
The idea is to train the model with 60 steps back from i and having 5 column target in i:
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
So my x-train (feed) and y-train (targets) are:
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)
Unfortunately, when fitting the model:
model.fit(X_train_splitted, y_train_splitted, epochs=3, batch_size=32)
I am getting an error:
Dimensions must be equal, but are 60 and 5 for '{{node mean_squared_error/SquaredDifference}} = SquaredDifference[T=DT_FLOAT](mean_squared_error/remove_squeezable_dimensions/Squeeze, IteratorGetNext:1)' with input shapes: [?,60], [?,5].
I understand that X_train_arr and y_train_arr need to be the same. BUT when testing with case below, everyting is fine:
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)
Idea of having print(y_train_arr.shape) #(2494, 5) is to be able to predict n-steps into the future, where each iteration of prediction generates new entire row of the data with 5 columns values.
Allright, after completing this tutorial i understood what should be done. Below is placed final code with comments:
#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