import numpy as np 
    import pandas as pd
    from matplotlib import pyplot as plt

    from keras.models import Sequential
    from keras.layers import Dense, Dropout

    import numpy as np
    from matplotlib import pyplot as plt

Data preparation

    # Dataset can be downloaded at

    data = pd.read_csv('Data/bike-sharing/hour.csv')
    # Feature engineering
    ohe_features = ['season', 'weathersit', 'mnth', 'hr', 'weekday']
    for feature in ohe_features:
        dummies = pd.get_dummies(data[feature], prefix=feature, drop_first=False)
        data = pd.concat([data, dummies], axis=1)

    drop_features = ['instant', 'dteday', 'season', 'weathersit', 'weekday', 'atemp', 'mnth', 'workingday', 'hr', 'casual', 'registered']
    data = data.drop(drop_features, axis=1)

    norm_features = ['cnt', 'temp', 'hum', 'windspeed']
    scaled_features = {}
    for feature in norm_features:
        mean, std = data[feature].mean(), data[feature].std()
        scaled_features[feature] = [mean, std]
        data.loc[:, feature] = (data[feature] - mean)/std

    # Save the final month for testing
    test_data = data[-31*24:]
    data = data[:-31*24]

    # Extract the target field
    target_fields = ['cnt']
    features, targets = data.drop(target_fields, axis=1), data[target_fields]
    test_features, test_targets = test_data.drop(target_fields, axis=1), test_data[target_fields]

    # Create a validation set (based on the last )
    X_train, y_train = features[:-30*24], targets[:-30*24]
    X_val, y_val = features[-30*24:], targets[-30*24:]

Without dropout

    model = Sequential()
    model.add(Dense(250, input_dim=X_train.shape[1], activation='relu'))
    model.add(Dense(150, activation='relu'))
    model.add(Dense(50, activation='relu'))
    model.add(Dense(25, activation='relu'))
    model.add(Dense(1, activation='linear'))

    # Compile model
    model.compile(loss='mse', optimizer='sgd', metrics=['mse'])

    n_epochs = 1000
    batch_size = 1024

    history =, y_train['cnt'], 
        validation_data=(X_val.values, y_val['cnt']), 
        batch_size=batch_size, epochs=n_epochs, verbose=0

    plt.plot(np.arange(len(history.history['loss'])), history.history['loss'], label='training')
    plt.plot(np.arange(len(history.history['val_loss'])), history.history['val_loss'], label='validation')
    plt.title('Overfit on Bike Sharing dataset')

The gap here means that the network learned very well the features from the training set but does not generalize to the validation data. This is what overfitting means.

    print('Minimum loss: ', min(history.history['val_loss']), 
     '\nAfter ', np.argmin(history.history['val_loss']), ' epochs')

    # Minimum loss:  0.129907280207 
    # After  980  epochs

Adding dropout

    model_drop = Sequential()
    model_drop.add(Dense(250, input_dim=X_train.shape[1], activation='relu'))
    model_drop.add(Dense(150, activation='relu'))
    model_drop.add(Dense(50, activation='relu'))
    model_drop.add(Dense(25, activation='relu'))
    model_drop.add(Dense(1, activation='linear'))

    # Compile model
    model_drop.compile(loss='mse', optimizer='sgd', metrics=['mse'])

    history_drop =, y_train['cnt'], 
        validation_data=(X_val.values, y_val['cnt']), 
        batch_size=batch_size, epochs=n_epochs, verbose=0

    plt.plot(np.arange(len(history_drop.history['loss'])), history_drop.history['loss'], label='training')
    plt.plot(np.arange(len(history_drop.history['val_loss'])), history_drop.history['val_loss'], label='validation')
    plt.title('Use dropout for Bike Sharing dataset')

There is no gap anymore between the training and validation. It means the network has not learned too much from the training but retained the essence.

    print('Minimum loss: ', min(history_drop.history['val_loss']), 
     '\nAfter ', np.argmin(history_drop.history['val_loss']), ' epochs')

    # Minimum loss:  0.126063346863 
    # After  998  epochs