def __encode__(self, **kwargs):
inputs = kwargs['inputs']
for k, v in inputs.items():
if inputs[k].shape == self.get_inputs_shape():
inputs[k] = tf.reshape(inputs[k],
(1, ) + self.get_inputs_shape())
inputs[k] = tf.cast(inputs[k], tf.float32)
kwargs['model'] = self.get_variable
kwargs['latents_shape'] = (self.batch_size, self.latents_dim)
encoded = self.encode_fn(**kwargs)
covariance_mean = infer_prior(latent_mean=encoded['z_latents'], \
regularize=False, lambda_d=self.lambda_d, lambda_od=self.lambda_od)
covariance_sigma = tf.exp(encoded['z_latents'] + epsilon)
latents_sigma = covariance_sigma
prior_distribution = tfp.distributions.Normal(loc=covariance_mean,
scale=covariance_sigma)
posterior_distribution = tfp.distributions.Normal(
loc=encoded['z_latents'], scale=latents_sigma)
bayesian_divergent = tfp.distributions.kl_divergence(
posterior_distribution, prior_distribution)
bayesian_divergent = tf.identity(bayesian_divergent,
name='bayesian_divergent')
return {**encoded, 'bayesian_divergent': bayesian_divergent}
def __encode__(self, **kwargs):
inputs = kwargs['inputs']
for k, v in inputs.items():
if inputs[k].shape == self.get_inputs_shape():
inputs[k] = tf.reshape(inputs[k], (1, ) + self.get_inputs_shape())
inputs[k] = tf.cast(inputs[k], tf.float32)
kwargs['model'] = self.get_variable
kwargs['latents_shape'] = (self.batch_size, self.latents_dim)
encoded = self.encode_fn(**kwargs)
_, covariance_regularizer = infer_prior(latent_mean=encoded['z_latents'], \
regularize=True, lambda_d=self.lambda_d, lambda_od=self.lambda_od)
return {**encoded, 'covariance_regularized': covariance_regularizer}