Может реализовывать ANN для прогнозирования временных рядов, но получать ошибки с помощью LSTM - PullRequest
Новая
       19

Может реализовывать ANN для прогнозирования временных рядов, но получать ошибки с помощью LSTM

0 голосов
/ 01 мая 2019

Я пытаюсь построить модель LSTM как другую модель для прогнозирования временных рядов.Мой набор данных pandas, dfTop50, состоит из 25 столбцов: Product_Code, M0, M1, ..., M23.Всего 50 строк.

Посмотрите пример того, как выглядит dfTop50 (фиктивные данные): enter image description here

Приведенный ниже код приводит к ошибке: 

melt = dfTop50.melt(id_vars='Product_Code', var_name='Month', value_name='Sales')
melt['Product_Code'] = melt['Product_Code'].astype(str)
melt['Month'] = melt['Month'].str.extract('(\d+)', expand=False).astype(int)
melt5 = melt.copy()
melt5['Last_Month_Sales'] = melt5.groupby(['Product_Code'])['Sales'].shift()
melt5['Last_Month_Diffs'] = melt5.groupby(['Product_Code'])['Last_Month_Sales'].diff()
melt5['Last-1_Month_Sales'] = melt5.groupby(['Product_Code'])['Sales'].shift(2)
melt5['Last-1_Month_Diffs'] = melt5.groupby(['Product_Code'])['Last-1_Month_Sales'].diff()
melt5['Last-2_Month_Sales'] = melt5.groupby(['Product_Code'])['Sales'].shift(3)
melt5['Last-2_Month_Diffs'] = melt5.groupby(['Product_Code'])['Last-2_Month_Sales'].diff()
melt5['Last-3_Month_Sales'] = melt5.groupby(['Product_Code'])['Sales'].shift(4)
melt5['Last-3_Month_Diffs'] = melt5.groupby(['Product_Code'])['Last-3_Month_Sales'].diff()
melt5['Last-4_Month_Sales'] = melt5.groupby(['Product_Code'])['Sales'].shift(5)
melt5['Last-4_Month_Diffs'] = melt5.groupby(['Product_Code'])['Last-4_Month_Sales'].diff()
melt5['Last-5_Month_Sales'] = melt5.groupby(['Product_Code'])['Sales'].shift(6)
melt5['Last-5_Month_Diffs'] = melt5.groupby(['Product_Code'])['Last-5_Month_Sales'].diff()
melt5['Last-6_Month_Sales'] = melt5.groupby(['Product_Code'])['Sales'].shift(7)
melt5['Last-6_Month_Diffs'] = melt5.groupby(['Product_Code'])['Last-6_Month_Sales'].diff()
melt5['Last-7_Month_Sales'] = melt5.groupby(['Product_Code'])['Sales'].shift(8)
melt5['Last-7_Month_Diffs'] = melt5.groupby(['Product_Code'])['Last-7_Month_Sales'].diff()
melt5['Last-8_Month_Sales'] = melt5.groupby(['Product_Code'])['Sales'].shift(9)
melt5['Last-8_Month_Diffs'] = melt5.groupby(['Product_Code'])['Last-8_Month_Sales'].diff()
melt5['Last-9_Month_Sales'] = melt5.groupby(['Product_Code'])['Sales'].shift(10)
melt5['Last-9_Month_Diffs'] = melt5.groupby(['Product_Code'])['Last-9_Month_Sales'].diff()
melt5['Last-10_Month_Sales'] = melt5.groupby(['Product_Code'])['Sales'].shift(11)
melt5['Last-10_Month_Diffs'] = melt5.groupby(['Product_Code'])['Last-10_Month_Sales'].diff()
melt5['Last-11_Month_Sales'] = melt5.groupby(['Product_Code'])['Sales'].shift(12)
melt5['Last-11_Month_Diffs'] = melt5.groupby(['Product_Code'])['Last-11_Month_Sales'].diff()
melt5 = melt5.dropna()
melt5 = melt5.drop(['Product_Code'], axis=1)

for month in range(18,19):
    train = melt5[melt5['Month'] < month]
    val = melt5[melt5['Month'] == month]
    xtr, xts = train.drop(['Sales'], axis=1), val.drop(['Sales'], axis=1)
    ytr, yts = train['Sales'].values, val['Sales'].values

    X_train = np.asmatrix(xtr)
    Y_train = np.asmatrix(ytr)
    X_test = np.asmatrix(xts)
    Y_test = np.asmatrix(yts)

    X_train_lmse = X_train.reshape(X_train.shape[0], X_train.shape[1], 1)
    X_test_lmse = X_test.reshape(X_test.shape[0], X_test.shape[1], 1)

    lstm_model = Sequential()
    lstm_model.add(LSTM(7, input_shape=(1, X_train_lmse.shape[1]), activation='relu', kernel_initializer='lecun_uniform', return_sequences=False))
    lstm_model.add(Dense(1))
    lstm_model.compile(loss='mean_absolute_error', optimizer='adam')
    early_stop = EarlyStopping(monitor='loss', patience=2, verbose=1)
    history_lstm_model = lstm_model.fit(X_train_lmse, Y_train, epochs=100, batch_size=1, verbose=1, shuffle=False, callbacks=[early_stop])
    nn_predictions_lstm = np.squeeze(model.predict(X_test_lmse))

    errorlstm = mean_absolute_percentage_error(Y_test, nn_predictions_lstm)
    print('Month %d - NN Error %.1f' % (month,errorlstm))

Я получаю сообщение об ошибке: "ожидал, что lstm_input будет иметь 3 измерения, но получил массив с формой (250,25).

Однако работает следующий код для ANN: 

for month in range(18,19):
    train = melt5[melt5['Month'] < month]
    val = melt5[melt5['Month'] == month]
    xtr, xts = train.drop(['Sales'], axis=1), val.drop(['Sales'], axis=1)
    ytr, yts = train['Sales'].values, val['Sales'].values

    EPOCHS = 100
    BATCH_SIZE = 10

    #Defining the 3 layered Neural Network
    def build_model():
        model = keras.Sequential([
        keras.layers.Dense(250, activation=tf.nn.softplus,
                       input_shape=(xtr.shape[1],)),
        keras.layers.Dense(250, activation=tf.nn.softplus),
        keras.layers.Dense(1)
        ])

        #Optimize the absolute error (prefer to squared error)
        model.compile(loss='mae',optimizer='adam', metrics=['mae'])
        return model

    model = build_model()
    model.summary()
    # Store training stats
    history = model.fit(xtr, ytr, epochs=EPOCHS, batch_size=BATCH_SIZE,
                    validation_split=0.0, verbose=0)
    nn_predictions0 = np.squeeze(model.predict(xts))
    error2 = mean_absolute_percentage_error(yts, nn_predictions0)
    print('Month %d - NN Error %.1f' % (month,error2))
...