Advanced Deep Learning with Keras

Chapter 8
log Pθ ( x | c) − DKL ( Qφ ( z | x, c) || Pθ ( z | x, c)
) = ⎡log P ( x | z, c) ⎤
⎣ θ ⎦
− DKL
( Qφ ( z | x, c) || Pθ
( z | c)
) (Equation 8.2.1)
z~
Q
Similar to VAEs, Equation 8.2.1 means that if we want to maximize the output
conditioned on c, P ( x | c)
, then the two loss terms must be minimized:
θ
• Reconstruction loss of the decoder given both the latent vector and the
condition.
• KL loss between the encoder given both the latent vector and the condition
and the prior distribution given the condition. Similar to a VAE, we typically
choose Pθ ( z | c) = P( z | c) = N ( 0, I ).
Listing 8.2.1, cvae-cnn-mnist-8.2.1.py shows us the Keras code of CVAE using
CNN layers. In the code that is highlighted showcases the changes made to support
CVAE:
# compute the number of labels
num_labels = len(np.unique(y_train))
# network parameters
input_shape = (image_size, image_size, 1)
label_shape = (num_labels, )
batch_size = 128
kernel_size = 3
filters = 16
latent_dim = 2
epochs = 30
# VAE model = encoder + decoder
# build encoder model
inputs = Input(shape=input_shape, name='encoder_input')
y_labels = Input(shape=label_shape, name='class_labels')
x = Dense(image_size * image_size)(y_labels)
x = Reshape((image_size, image_size, 1))(x)
x = keras.layers.concatenate([inputs, x])
for i in range(2):
filters *= 2
x = Conv2D(filters=filters,
kernel_size=kernel_size,
activation='relu',
strides=2,
padding='same')(x)
# shape info needed to build decoder model
[ 255 ]