Examples and Usage of TimeWarPy

Preprocessing Univariate Data for Recurrent Neural Network Training

With less than 30 lines, you can use TensorFlow and TimeWarPY to create a fully trained model for time series forecasting. A comparison of the complexity decrease can be seen here in the TensorFlow docs.

# load libraries
import tensorflow as tf
import numpy as np
from sklearn.model_selection import train_test_split
from sklearn.preprocessing import MinMaxScaler
from timewarpy import core, datasets

# load and preprocess
df = datasets.load_energy_data()
TSprocessor = core.UnivariateTS(1000, 100, scaler=MinMaxScaler)
X, y = TSprocessor.fit_transform(df, 'Appliances')
X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.33, random_state=42)
print(f'Original dataframe shape: {df.shape}')
print(f'X training vector shape: {X_train.shape}')
print(f'y training vector shape: {y_train.shape}')

# train small tensorflow model
model = tf.keras.models.Sequential()
model.add(tf.keras.layers.LSTM(10, activation='tanh', recurrent_activation='sigmoid', input_shape=X_train[0].shape))
model.add(tf.keras.layers.Dense(100))
model.compile(optimizer='Adam', loss='mean_squared_error',)
history = model.fit(X_train, y_train, epochs=2, batch_size=100,)

# make predictions
y_pred = model.predict(X_test)
mae = np.mean(np.abs(TSprocessor.inverse_transform(y_test - y_pred)))
print(f'y prediction vector shape: {y_pred.shape}')
print(f'Model Mean Absolute Error: {mae:.2f}')

Original dataframe shape: (19735, 29)
X training vector shape: (12486, 1000, 1)
y training vector shape: (12486, 100)
Epoch 1/2
125/125 [==============================] - 62s 463ms/step - loss: 0.0103
Epoch 2/2
125/125 [==============================] - 52s 416ms/step - loss: 0.0085
y prediction vector shape: (6150, 100)
Model Mean Absolute Error: 56.82

Preprocessing Multivariate Data for Recurrent Neural Network Training

Similar to the univariate example above, you also might want to use multiple variables as features for your time series predictions. You also might want to predict more than one variable based on the given problem. The example below show how to use three features to predict a timeseries of 2 variables.

# load libraries
import tensorflow as tf
import numpy as np
from sklearn.model_selection import train_test_split
from sklearn.preprocessing import MinMaxScaler
from timewarpy import core, datasets

# load and preprocess
df = datasets.load_energy_data()
TSprocessor = core.MultivariateTS(1000, 100, scaler=MinMaxScaler)
X, y = TSprocessor.fit_transform(df, train_columns=['Appliances', 'T1', 'RH_1'], pred_columns=['Appliances', 'T1'])
X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.33, random_state=42)
print(f'Original dataframe shape: {df.shape}')
print(f'X training vector shape: {X_train.shape}')
print(f'y training vector shape: {y_train.shape}')

# train small tensorflow model
model = tf.keras.models.Sequential()
model.add(tf.keras.layers.LSTM(10, activation='tanh', recurrent_activation='sigmoid', input_shape=X_train[0].shape))
model.add(tf.keras.layers.Dense(200))
model.add(tf.keras.layers.Reshape((100, 2), input_shape=(200,)))
model.compile(optimizer='Adam', loss='mean_squared_error',)
history = model.fit(X_train, y_train, epochs=2, batch_size=100,)

# make predictions
y_pred = model.predict(X_test)
print(f'y prediction vector shape: {y_pred.shape}')

Original dataframe shape: (19735, 29)
X training vector shape: (12486, 1000, 3)
y training vector shape: (12486, 100, 2)
Epoch 1/2
125/125 [==============================] - 68s 523ms/step - loss: 0.0610
Epoch 2/2
125/125 [==============================] - 67s 537ms/step - loss: 0.0112
y prediction vector shape: (6150, 100, 2)