deep-learning-with-python-notebooks
deep-learning-with-python-notebooks copied to clipboard
fchollet
2nd edition Chapter 12, VAE latent space sampler differences with 1st edition
In the 2nd edition notebook of chapter 12 part 4, the latent space sampling layer is defined as the following class:
class Sampler(layers.Layer):
def call(self, z_mean, z_log_var):
batch_size = tf.shape(z_mean)[0]
z_size = tf.shape(z_mean)[1]
epsilon = tf.random.normal(shape=(batch_size, z_size))
return z_mean + tf.exp(0.5 * z_log_var) * epsilon
In the 1st edition, it's defined as the following fucntion:
def sampling(args):
z_mean, z_log_var = args
epsilon = K.random_normal(shape=(K.shape(z_mean)[0], 2),
mean=0., stddev=1.)
return z_mean + K.exp(z_log_var) * epsilon
Most changes come from following best practices and API changes, but I do have the folowing doubt: Why is z_log_var multiplied by 0.5 in the 2nd edition class? Is it because a difference between tf.exp() and tf.keras.backend.exp()? Is it becasue reducing it by half yields better results? Any other reason?
