keras_average.py

from numpy.random import seed
from pandas import read_csv, DataFrame
from sklearn.preprocessing import scale
from keras.layers.convolutional import Conv1D, MaxPooling1D
from keras.layers.merge import average
from keras.layers import Input, Dense, Flatten, Reshape, Dropout, SpatialDropout1D
from keras.models import Model
from keras.optimizers import SGD
from keras.utils import plot_model

df = read_csv("credit_count.txt")
Y = df[df.CARDHLDR == 1].DEFAULT
X = scale(df[df.CARDHLDR == 1].iloc[:, 2:12])
D = 0.2
S = 1

seed(S)
### INPUT DATA
inputs = Input(shape = (X.shape[1],))
### DEFINE A MULTILAYER PERCEPTRON NETWORK
mlp_net = Dense(64, activation = 'relu', kernel_initializer = 'he_uniform')(inputs)
mlp_net = Dropout(rate = D, seed = S)(mlp_net)
mlp_net = Dense(64, activation = 'relu', kernel_initializer = 'he_uniform')(mlp_net)
mlp_net = Dropout(rate = D, seed = S)(mlp_net)
mlp_out = Dense(1, activation = 'sigmoid')(mlp_net)
mlp_mdl = Model(inputs = inputs, outputs = mlp_out)
### DEFINE A CONVOLUTIONAL NETWORK 
cnv_net = Reshape((X.shape[1], 1))(inputs)
cnv_net = Conv1D(32, 4, activation = 'relu', padding = "same", kernel_initializer = 'he_uniform')(cnv_net)
cnv_net = MaxPooling1D(2)(cnv_net)
cnv_net = SpatialDropout1D(D)(cnv_net)
cnv_net = Flatten()(cnv_net)
cnv_out = Dense(1, activation = 'sigmoid')(cnv_net)
cnv_mdl = Model(inputs = inputs, outputs = cnv_out)
### COMBINE MLP AND CNV
con_out = average([mlp_out, cnv_out])
con_mdl = Model(inputs = inputs, outputs = con_out)
sgd = SGD(lr = 0.1, momentum = 0.9)
con_mdl.compile(optimizer = sgd, loss = 'binary_crossentropy',  metrics = ['binary_accuracy'])
con_mdl.fit(X, Y, batch_size = 2000, epochs = 50, verbose = 0)
plot_model(con_mdl, to_file = 'model.png', show_shapes = True, show_layer_names = True)