def discriminator(input_shape,
optimizer,
d_activation,
d_batchnorm,
dropout=0.1,
width_modifier=0.5):
D = Sequential()
# Input layer
input_layer_width = input_shape
D.add(Dense(input_layer_width, input_shape=(input_layer_width, )))
D.add(d_activation)
if d_batchnorm:
D.add(BatchNormalization())
D.add(Dropout(dropout))
# First Hidden Layer
first_layer_width = int(input_shape * width_modifier)
D.add(Dense(first_layer_width))
D.add(d_activation)
if d_batchnorm:
D.add(BatchNormalization())
D.add(Dropout(dropout))
# Second Hidden Layer
second_layer_width = int(input_shape * width_modifier)
D.add(Dense(second_layer_width))
D.add(d_activation)
if d_batchnorm:
D.add(BatchNormalization())
D.add(Dropout(dropout))
# Third Hidden Layer
third_layer_width = int(input_shape * width_modifier)
D.add(Dense(third_layer_width))
D.add(d_activation)
if d_batchnorm:
D.add(BatchNormalization())
D.add(Dropout(dropout))
# Output
D.add(Dense(1, activation='sigmoid'))
D._name = 'discriminator'
D.compile(optimizer=optimizer,
loss=label_smooth_sigmoid,
metrics=['accuracy'])
D.summary()
return D
def generator(num_inputs, output_shape, optimizer, g_activation, g_batchnorm, generator_activation, dropout = 0.1, width_modifier = 0.5):
G = Sequential()
# Input layer
input_layer_width = num_inputs
G.add(Dense(input_layer_width, input_shape = (input_layer_width,)))
G.add(g_activation)
if g_batchnorm:
G.add(BatchNormalization())
G.add(Dropout(dropout))
# First Hidden Layer
first_layer_width = int(num_inputs * width_modifier)
G.add(Dense(first_layer_width))
G.add(g_activation)
if g_batchnorm:
G.add(BatchNormalization())
G.add(Dropout(dropout))
# Second Hidden Layer
second_layer_width = int(num_inputs * width_modifier)
G.add(Dense(second_layer_width))
G.add(g_activation)
if g_batchnorm:
G.add(BatchNormalization())
G.add(Dropout(dropout))
# Third Hidden Layer
third_layer_width = int(num_inputs * width_modifier)
G.add(Dense(third_layer_width))
G.add(g_activation)
if g_batchnorm:
G.add(BatchNormalization())
G.add(Dropout(dropout))
# Output layer
G.add(Dense(output_shape, activation = 'sigmoid'))
G._name = 'generator'
G.compile(optimizer = optimizer, loss = 'categorical_crossentropy')
G.summary()
return G
